parallel lstm-based regional integrated energy system...

14
Research Article Parallel LSTM-Based Regional Integrated Energy System Multienergy Source-Load Information Interactive Energy Prediction Bo Wang , 1 Liming Zhang, 1 Hengrui Ma, 2 Hongxia Wang , 1 and Shaohua Wan 3 1 School of Electrical Engineering, Wuhan University, Wuhan 430070, China 2 Tus-Institute for Renewable Energy, Qinghai University, Xining, China 3 School of Information and Safety Engineering, Zhongnan University of Economics and Law, Wuhan 430073, China Correspondence should be addressed to Shaohua Wan; [email protected] Received 16 July 2019; Accepted 17 October 2019; Published 22 November 2019 Guest Editor: Xiaoqing Bai Copyright © 2019 Bo Wang et al. is is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. e multienergy interaction characteristic of regional integrated energy systems can greatly improve the efficiency of energy utilization. is paper proposes an energy prediction strategy for multienergy information interaction in regional integrated energy systems from the perspective of horizontal interaction and vertical interaction. Firstly, the multienergy information coupling correlation of the regional integrated energy system is analyzed, and the horizontal interaction and vertical interaction mode are proposed. en, based on the long short-term memory depth neural network time series prediction, parallel long short- term memory multitask learning model is established to achieve horizontal interaction among multienergy systems and based on user-driven behavioral data to achieve vertical interaction between source and load. Finally, uncertain resources composed of wind power, photovoltaic, and various loads on both sides of source and load integrated energy prediction are achieved. e simulation results of the measured data show that the interactive parallel prediction method proposed in this article can effectively improve the prediction effect of each subtask. 1. Introduction With the depletion of fossil energy, the contradiction be- tween energy demand growth and energy shortage in the process of social and economic development, energy utili- zation, and environmental protection is becoming more and more serious. Under such background, the consumption system of traditional energy production with fossil energy as the core has been difficult to sustain, so building a multi- energy complementary and optimized energy supply and demand system and guiding the overall transformation of the energy industry have become the top priority of China’s energy sector development [1]. On the basis of energy system source-network-load-storage vertical optimization, the multienergy supply systems are coordinated and optimized in a horizontal direction by the multienergy coupling re- lationships to realize integrated energy system for energy cascade utilization and multienergy coordinated scheduling, which will play a key role in the abovementioned energy revolution [2]. On the one hand, it will improve energy efficiency through comprehensive development and utili- zation of energy, and on the other hand, it will increase the potential energy penetration rate by converting electricity into heat, cold, natural gas, electric vehicle energy storage, etc. e extension of IES in end users is called regional integrated energy system (RIES). In the RIES with in- termittent renewable energy such as wind power/photo- voltaic, due to the coupling and interconnection of various energy subsystems such as power system, thermal system, natural gas system, and transportation system, the energy consumption data are scattered and multidimensional of the versatile load such as electric load, heat load, and gas load on the user side. Such huge amount of data information is of great value in situational awareness, especially in new energy generation and load forecasting [3]. Zhang et al. [4] used Hindawi Complexity Volume 2019, Article ID 7414318, 13 pages https://doi.org/10.1155/2019/7414318

Upload: others

Post on 18-Oct-2020

5 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Parallel LSTM-Based Regional Integrated Energy System ...downloads.hindawi.com/journals/complexity/2019/7414318.pdf · Characteristics of Regional Integrated Energy System “Information

Research ArticleParallel LSTM-Based Regional Integrated Energy SystemMultienergy Source-Load Information InteractiveEnergy Prediction

Bo Wang 1 Liming Zhang1 Hengrui Ma2 Hongxia Wang 1 and Shaohua Wan 3

1School of Electrical Engineering Wuhan University Wuhan 430070 China2Tus-Institute for Renewable Energy Qinghai University Xining China3School of Information and Safety Engineering Zhongnan University of Economics and Law Wuhan 430073 China

Correspondence should be addressed to Shaohua Wan shaohuawanieeeorg

Received 16 July 2019 Accepted 17 October 2019 Published 22 November 2019

Guest Editor Xiaoqing Bai

Copyright copy 2019 Bo Wang et al is is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

e multienergy interaction characteristic of regional integrated energy systems can greatly improve the eciency of energyutilization is paper proposes an energy prediction strategy for multienergy information interaction in regional integratedenergy systems from the perspective of horizontal interaction and vertical interaction Firstly the multienergy informationcoupling correlation of the regional integrated energy system is analyzed and the horizontal interaction and vertical interactionmode are proposeden based on the long short-term memory depth neural network time series prediction parallel long short-term memory multitask learning model is established to achieve horizontal interaction among multienergy systems and based onuser-driven behavioral data to achieve vertical interaction between source and load Finally uncertain resources composed ofwind power photovoltaic and various loads on both sides of source and load integrated energy prediction are achieved esimulation results of the measured data show that the interactive parallel prediction method proposed in this article can eectivelyimprove the prediction eect of each subtask

1 Introduction

With the depletion of fossil energy the contradiction be-tween energy demand growth and energy shortage in theprocess of social and economic development energy utili-zation and environmental protection is becoming more andmore serious Under such background the consumptionsystem of traditional energy production with fossil energy asthe core has been dicult to sustain so building a multi-energy complementary and optimized energy supply anddemand system and guiding the overall transformation ofthe energy industry have become the top priority of Chinarsquosenergy sector development [1] On the basis of energy systemsource-network-load-storage vertical optimization themultienergy supply systems are coordinated and optimizedin a horizontal direction by the multienergy coupling re-lationships to realize integrated energy system for energycascade utilization and multienergy coordinated scheduling

which will play a key role in the abovementioned energyrevolution [2] On the one hand it will improve energyeciency through comprehensive development and utili-zation of energy and on the other hand it will increase thepotential energy penetration rate by converting electricityinto heat cold natural gas electric vehicle energy storageetc

e extension of IES in end users is called regionalintegrated energy system (RIES) In the RIES with in-termittent renewable energy such as wind powerphoto-voltaic due to the coupling and interconnection of variousenergy subsystems such as power system thermal systemnatural gas system and transportation system the energyconsumption data are scattered and multidimensional of theversatile load such as electric load heat load and gas load onthe user side Such huge amount of data information is ofgreat value in situational awareness especially in new energygeneration and load forecasting [3] Zhang et al [4] used

HindawiComplexityVolume 2019 Article ID 7414318 13 pageshttpsdoiorg10115520197414318

data analysis and prediction algorithms to verify the cor-relation between energy consumption behavior and dis-tributed PV output and input the prediction results into thepredicted energy management system to optimize load-sideenergy consumption and balance the systemrsquos energy supplyand load demand Deng et al [5] proposed the energy 50concept based on cyber-physical-social systems (CPSS) andanalyzed the social behavior characteristics of load-sideenergy consumption in smart home energy systems bycollecting a variety of operational information in the actualsystem by using the ldquodata driven + physical modelrdquo approachto form a new energy forecasting agent to achieve optimaloperational control [6ndash8] In addition in the energy pre-diction of RIES deep learning prediction technology hasattracted the attention of many scholars at home and abroadWang et al [9] have made extensive analysis and discussionon renewable energy forecasting methods based on deeplearning discussed the current problems facing this di-rection and prospected the future development directionWang et al [10] proposed a wind speed prediction modelbased on deep confidence network and achieved goodprediction results Chen et al [11] proposed a new two-layernonlinear combination short-term wind speed predictionmethod (EEL-ELM) In the first layer the extreme learningmachine (ELM) Elman neural network (ENN) and longshort-term memory (LSTM) neural network are used topredict wind speed and three prediction results are ob-tained +en the nonlinear aggregation mechanism of ex-treme learning machine (ELM) is used to alleviate theinherent weakness of single method and linear combinationLiu et al [12] combined variational mode decomposition(VMD) singular spectrum analysis (SSA) long short-termmemory (LSTM) network and extreme learning machine(LEM) and proposed a new multistep wind speed predictionmodel +e simulation results show that the method is moreeffective and robust in extracting trend information On thebasis of multiple long short-term memory (LSTM) learningthe literature [13] combines the extreme value optimizationalgorithm and the support vector machine model to in-tegrate the LSTM layer prediction results and predicts thewind speed in a short time In literature [14] the LSTM deepneural network algorithm is used and the userrsquos energy dataof various types of equipment are used to predict the res-idential load on the load side in a short term which provesthe data value of the load energy information in the fore-casting field Multitask learning has been studied for about20 years [15ndash17] At present multitask learning methods canbe roughly summarized into two categories One is to sharethe same parameters between different tasks and the other isto mine the shared data features hidden between differenttasks [18] Argyriou et al [19] detailed the four multitaskinglearning methods of feature selection kernel selectionadaptive pooling and graphical model structure Jebara [20]proposed a typical multitask model for mining commonfeatures between multitasks A framework for multitaskfeature learning is given in this paper which becomes thebasis for many multitask learning references [21 22] In-spired by the above literature this paper argues that it isnecessary to explore the value of multienergy coupling

information and load-side user behavior characteristics forenergy system operation and use the information correlationcharacteristics between multiple energy sources to expandthe concept of multienergy interaction to the forecastingstage +e ldquoinformation interactionrdquo strategy is proposedand explored the data information implied in each link ofRIES by combining deep learning technology and the en-ergy utilization efficiency is maximized through multienergyinteraction Based on the existing research this papercomprehensively considers the horizontal and vertical in-formation interaction characteristics between multiple en-ergy loads in RIES and proposes a parallel energy LSTMbased on multienergy-load information interactive short-term energy prediction strategy Firstly the RIES multi-energy-charge coupling correlation is analyzed and thehorizontal and vertical ldquoinformation interactionrdquo modes ofmultienergy load are proposed +en from the perspectiveof information interaction based on the LSTM deep neuralnetwork time series prediction the parallel LSTM multitasklearning model is established to realize the horizontal in-teraction between multienergy systems At the same timebased on the userrsquos energy behavior data driving the verticalinteraction between the source and the load is realized andthe uncertain sources are used for integrated energy pre-diction of the wind power photovoltaic and various loadson both sides of the source and the load Finally based on themeasured data of a regionrsquos RIES the prediction model istrained and verified +e simulation results show that theproposed information interactive parallel prediction methodcan effectively improve the prediction effect of each subtask

2 Regional Integrated Energy System

21 Basic Structure of Regional Integrated Energy System+e regional integrated energy system takes the electricity asthe core is based on smart grid and is led by clean energy Ituses advanced information communication and energyconversion technologies to organically connect the links ofproducing transporting storing and consuming in themultienergy systems such as electricity gas and heat so as toachieve optimal allocation of energy multienergy couplingand complement [23] RIES structurally includes regionalenergy production input network multienergy couplingnetwork and regional internal load-side energy outputnetwork It can fully absorb all kinds of renewable energysuch as wind power and photovoltaic and access varioustypes of loads such as electricity gas heat and cold toachieve efficient energy allocation At the same time its toplayer has an integrated information communication supportsystem to achieve the interactive fusion of energy flow andinformation flow +e basic architecture of the RIES de-scribed in this paper is shown in Figure 1

22 Characteristics of Regional Integrated Energy SystemldquoInformation Interconnectionrdquo Under the support of theenergy Internet and multienergy key technologies charac-terized by ldquoopenrdquo and ldquointerconnectedrdquo the coupling in-teraction of multienergy-network-load-storage in RIES is

2 Complexity

more tight [24] (1) From the perspective of horizontalinterconnection RIES will integrate the energy subsystemssuch as coldheatelectricgastraffic to construct a mul-tienergy flow system +e energy is converted and flowedbetween the systems through the coupling componentsand the various energy forms complement each other +eenergy resources have to be optimally configured in a widearea that is to achieve the ldquohorizontal information in-terconnectionrdquo of the energy Internet (2) From the per-spective of vertical interconnection based on thetraditional energy supply network RIES will realize thereal-time sharing of information resources in the source-network-load-storage through Internet technology +isresource sharing mechanism has two-way conductioncharacteristics +e user has information bidirectionalconduction capability between the information controlsystem the energy supply module and the multienergyintelligent delivery module +e interactive sharing ofenergy information can also be realized between users thatis to realize the ldquovertical information interconnectionrdquo ofthe energy Internet

3 RIES Multienergy CouplingCorrelation Analysis

31 RIES Horizontal Multienergy Information Coupling+e horizontal multienergy information coupling corre-lation of RIES is represented by the correlation betweendifferent energy flow subsystems such as electricity gasand heat

From the perspective of coupling the composition of theRIES can be divided into a coupling unit and a noncouplingunit As shown in Figure 2 the coupling unit connectsdifferent energy flow subsystems and consumes one or moreenergy sources to generate other energy sources Typicalcoupling forms are as follows cold and heat electricity

supply units consume coal or natural gas to generate elec-tricity heat and cold +e heat pump consumes electricalenergy to generate thermal energy the electrical hydrogenconsumes electrical energy to generate hydrogen and thecoupling unit causes the energy flow subsystems to interact+e uncoupled unit includes the network and equipmentinside each energy flow subsystem and other uncoupledboundary conditions such as source and load +e uncou-pled unit also affects other systems through the coupling unit[25]+erefore different energy flow subsystems have strongcoupling between them and there is a large amount ofhorizontal multienergy coupling information existing in themultienergy system data resulting in great correlation be-tween various units within the system

32 RIES Vertical Source-Load Information Coupling +evertical source-load information coupling correlation ofRIES is represented by the correlation between the multi-energy supply of the source and the userrsquos energy behavior of

WT

Input side

Distributionnetwork

Natural gas WindSolar

Output side Cold load Electric

loadHeat load

PV

Refrigeration

CCHPGB

ES

HS

Coupling side

Informationcommunicationsupport system

Calculationdata storage

layer

Informationcorrespondence

network

Acquisitionmonitoring

layer

Analysisdecision level

Lateral information interaction between multiplecapabilities

Vert

ical

info

rmat

ion

inte

ract

ion

betw

een

sour

ce lo

ads

sum

sum

Figure 1 Basic framework of RIES

Gas

Gasturbine

CHP

Powerto gas

Electric storage

Heatpump

Gasboiler

HeatCirculating pump

Heat storage

Gas storage

Electric

Figure 2 Transverse multicoupling schematic diagram

Complexity 3

the load Different energy behaviors correspond to differentload characteristics [26] For example the lighting behaviorin RIES is a typical user energy behavior +rough advancedmeasurement devices the userrsquos lighting behavior can bereflected as a lighting load curve

From the perspective of user energy behavior there aremultiple coupling relationships between multiple energyloads in RIES Taking the userrsquos lighting behavior as anexample if the cloud suddenly covers a certain area andreduces the ambient light intensity it will affect the userrsquoslighting behavior people will increase the lighting equip-ment usage to meet the lighting demand and the lightingload will increase At the same time such cloud movementswill also have an impact on photovoltaic power generationcloud clusters will block the photovoltaic panels which willaffect the irradiance reduction on the photovoltaic panelsand affect the photovoltaic power generation In order toprove the characteristics of this behavioral feature this papertakes the data of three days from January 24 to 26 2016 ofRIES in a certain area of Tianjin for quantitative analysis+ephotovoltaic power output curve and the user lighting loadcurve are shown in Figure 3

+e upper part of Figure 3 is the curve trend of thephotovoltaic power output and the lower part is the curvetrend of the userrsquos lighting load (1) Both the userrsquos lightingload and the photovoltaic power output have a day and nightcycle +e distribution of them is mainly concentrated in thedaytime +e peak of the userrsquos lighting load is generallydistributed between 18 00 and 20 00 and the peak outputof the photovoltaic power supply is generally distributedbetween 12 00 and 14 00 at noon (2) Taking the three-dayshort-term data for comparison analysis the overall outputof photovoltaic power generation on the third day wassignificantly larger than the previous two days Accordinglythe lighting load on the third day between 10 00 and 16 00is reduced compared to the previous two days and it isreduced to the low of the day It shows that when the outputof photovoltaic power generation increases it also affects theuserrsquos energy consumption behavior when the ambient lightintensity is large which means that the userrsquos demand forlighting energy is reduced (3) Take the ultra-short-termdata analysis from 12 00 to 16 00 on the first day andthe photovoltaic power output reaches the peak of one day at13 00 and the lighting load is also reduced to low of thedaytime at this moment During the period from 13 00 to14 00 a sudden downward climb occurs and the lightingload increases at the corresponding time +is ldquomutationrdquophenomenon should be caused by the cloudy weather af-fected by the cloud From the above analysis it can be foundthat there is a strong information coupling correlation be-tween the ldquosourcerdquo photovoltaic power output and theldquocharge endrdquo userrsquos lighting load

33 RIES Multienergy Information Interaction ModeRIES has both horizontal multienergy information couplingcorrelation characteristics and vertical source-load in-formation coupling correlation characteristics From theperspective of information interaction energy management

there are also two horizontal and vertical information in-teraction modes

+e horizontal information interaction mode is based onmultienergy conversion information sharing On the basis ofobtaining the measurement data of the multienergy flowsubsystem the nonlinear artificial intelligence deep learningtechnology is used to process the multienergy data in par-allel and it effectively utilizes the complex shared in-formation of energy conversion identifies the abstractfeatures of the training data and improves the precision ofsubtask prediction of the energy management

+e vertical information interaction mode is based onuser behavior information sharing On the basis of obtainingmulticategory data on both sides of the source and the loadthe deep learning method is used to dig the user-side be-havior characteristics and the associated user behaviorinformation data are taken as input to guide the source-sideenergy management prediction and improve the predictionaccuracy of the energy supply end

4 Multitask Learning Mechanism for DeepLearning Energy Prediction

41 LSTM Recurrent Neural Network

411 Principle of LSTM Cyclic Neural Network +e trainingobjects of the source-charged uncertain resources haveobvious timing characteristics and the sequence data havestrong correlation before and after From the perspective oflearning and training recurrent neural network (RNN) canprocess sequences with temporal correlation of arbitrarylength by using neurons with self-feedback which has ob-vious advantages in processing sequence data [27] +estructure of the RNN is shown in Figure 4(a) +e LSTMnetwork is a variant of RNN that overcomes the traditionalRNN gradient demise problem [28] +e hidden layer is nolonger a simple neural unit but an LSTM unit with a uniquememory mode Each LSTM unit contains a state cell thatdescribes the current state of the LSTM unit [29] c repre-sents the current state quantity of the LSTM unit and h

represents the current output of the LSTM unitLSTM reads and modifies state units by controlling the

forget gates input gates and output gates [30] which are

0

ndash100

250

50

100

150

200

ndash50

ndash200

ndash100

0

100

200

300

P (k

W)

10 20 30 40 50 60 700t (h)

Figure 3 +e relationship between the photovoltaic power curveand the lighting load curve of the user

4 Complexity

typically formed by sigmoid or tanh functions and Hada-mard product operation As shown in Figure 4(b) the forgetgate determines ctminus 1 of the previous moment and how manycomponents of the cell state ct remain at the current timeand the input gate determines how many components of thenetwork input are saved to the cell state ct at the current timext and howmany components of the output gate control cellstate ct are output to the current value ht of the LSTM

+e calculation formula between the variables of theLSTM unit is as follows

ft σ Wfxxi + Wfhhtminus 1 + bf1113872 1113873

it σ Wixxi + Wihhtminus 1 + bi( 1113857

gt ϕ Wgxxt + Wghhtminus 1 + bg1113872 1113873

ct ft ⊙ ctminus 1 + it ⊙gt

ot σ Woxxt + Wohhtminus 1 + bo( 1113857

ht ot ⊙ϕ ct( 1113857

(1)

In the formulaWfxWfhWixWihWgxWghWox andWoh are the weight matrices corresponding to the input of thenetwork activation function bf bi bg and bo are the offsetvectors ⊙ represents the multiplication of the Hadamardproduct matrix σ represents the sigmoid activation functionand ϕ represents the tanh activation function

412 LSTMDeep Learning Training Algorithm +e trainingof LSTM network adopts BP-based backpropagationthrough time (BPTT) [31] +e algorithm steps are as fol-lows (1) Calculate the output value of each neuron forwardFor the LSTMunit the values of the six vectors such asft itgt ct ot and ht are calculated (2) Calculate the error termof each neuron in reverse Similar to traditional RNN thebackpropagation of the LSTM error term consists of twolevels one is at the spatial level and the error term ispropagated to the upper layer of the network the other is atthe time level which propagates back in time For examplefrom the current time of t the error at each moment iscalculated (3) Calculate the gradient of each weightaccording to the corresponding error term and update thenetwork weight parameter (4) Jump to the first step andcontinue to perform the first 3 steps until the network erroris less than the given value

42 LSTM Network Multitask Learning MechanismMultitask learning (MTL) as an inductive migrationmechanism avoids the underfitting of model training byusing the correlation between multiple tasks digging thehidden common features and improving the generalizationlearning ability of the model [32] RIES source wind andlight prediction and load-end electricity gas heat and coldload prediction can be regarded as different tasks in mul-titask learning [33] Multiple tasks share some common dataor model structures that is there is a certain correlationbetween the multienergy prediction tasks +e multitasklearning prediction model based on the LSTM networkproposed in this paper shares the hidden layer nodes amongall prediction tasks through the hard parameter sharingmethod and trains the training data corresponding to themultienergy prediction tasks at the same time consideringthe correlation information between tasks to improve thelearning ability of all predictive tasks [34] A comparison ofsingle-task and multitask learning based on the LSTMnetwork is shown in Figure 5 +e multisource historicaldata in the figure is a collection of five types of historical dataincluding wind power photovoltaic electric load heat loadand cold load

5 Parallel LSTM-Based RIES InformationInteractive Energy Prediction Method

51 RIES Multienergy Prediction Model Design RIES hasuncertainties at both the source and the load end so accurateprediction of uncertain resources is the basis for imple-menting RIES energy optimization management andscheduling [35] +e source mainly includes wind powerprediction and photovoltaic power prediction and the loadmainly includes electric load prediction thermal load pre-diction and cold load prediction

As described in the second section of this paper there areintricate mutual coupling relationships between multipleenergy loads of RIES and the prediction tasks of uncertainresources on both sides of the source and load have strongcorrelation It can effectively improve the overall predictionaccuracy of the RIES uncertain resources by learning thiscoupling relationship through joint prediction training [36]From the perspective of RIES multienergy-load informationinteraction this paper proposes the concept of ldquoparallel

xtndash1 xt xt+1

htndash2

ctndash2

hh h

cc c

htndash1 ht ht+1

ctndash1 ct ct+1

(a)

tanh

tanh

xt

ht

ctctndash1

Forgetgate

ft it ot

ct

htndash1

Inputgate

Outputgate

σ σ σ

(b)

Figure 4 Schematic diagram of RNN and LSTM (a) +e structure of the RNN (b) +e structure of the LSTM cell

Complexity 5

predictionrdquo and utilizes the strong nonlinear learning abilityand historical memory advantages of LSTM deep learningnetwork to conduct integrated prediction of uncertain re-sources at both source and load As shown in Figure 6 thehorizontal information interaction based on multienergyconversion information sharingmainly utilizes themultitasklearning parameter sharing mechanism of the parallel LSTMprediction network the vertical information interactionbased on the user behavior information sharing mainly usesthe user behavior data to guide source-load multitask pre-diction as an additional unified input information Para-metric sharing is weight sharing which takes advantage ofthe correlation between subtasks Input sharing refers todata sharing Different subtasks share historical power datasuch as wind power photovoltaic power power loadthermal load and cooling load

52 Data Preprocessing +e training samples select the windpower generation time series on the source side the photo-voltaic power generation time series the charge side on the loadside the heat load the cold load time series and the illumi-nation load series data Considering the input and output rangeof the nonlinear activation function in the model in order toavoid neuron saturation the data are normalized and mappedbetween [minus 1 1] as shown in the following equation

xprime x minus xmax + xmin( 11138572

xmax minus xmin( 11138572 (2)

In the equation x is the actual input or output data andxmax and xmin are the maximum and minimum values of thevariable respectively

53 Task Construction +e essence of using the multitasklearning method for multienergy-load integration predictionlies in the close correlation of each prediction subtask [37]+e traditional single-task prediction method uses a singledimension of historical observations to predict its futureoutput and the model output is a single attribute +emultitask learning integration prediction uses the multidi-mensional data on both sides of the source and the load as a

common input to effectively utilize the sample On the otherhand through the shared hidden layer node divisionmethodthe five attributes of wind photovoltaic power generationelectricity heat and cold load are simultaneously predictedto play the role of inductive bias and improve predictionaccuracy

54 Model Training +e LSTM-based source-charge in-tegrated prediction model mainly needs to determine thefollowing parameters input layer dimension input layertime step hidden layer number each hidden layer di-mension and output layer dimension [38]

+e time step of the input layer is generally the length ofthe variable time series used for the source-load integrationprediction According to the periodic characteristics of theprediction object sequence the input layer time step is set to24 that is the historical data whose resolution ratio is 1 hduring the first seven days before the input is used forforecasting and it takes up one input node per day+e inputlayer dimension is the number of nodes +e input layernode of this paper contains six objects wind power gen-eration time series photovoltaic power generation timeseries charge side load heat load cold load time series andlighting load series which constitutes totally 42 nodes so theinput layer dimension is set to 42 Since this paper simul-taneously predicts six attributes of wind power photovol-taic electricity heat and cold load the output layer has 5nodes that is the output layer dimension is set to 5 In thispaper based on experience and multiple trial and com-parison the number of hidden layers is set to 2 and thedimension of the hidden layer is set to 30 +e selection ofthe activation function is as shown in Figure 4(a) thebatch_size is set to 42 the numbers of iterations of an epochis set to 8784 the learning rate is set to 080 and the learningdelay rate is set to 005 +e network training function usesthe BPTT algorithm mentioned above

55 Evaluation Indicators Multienergy-load integrationforecasting simultaneously trains multiple forecastingtasks To comprehensively measure the forecasting effect itis necessary to evaluate the integrated forecasting accuracyas a whole [39] In this paper the average accuracyevaluation index based on weight coefficient is proposedFor the importance degree of different energy on the twosides of the RIES the corresponding weight coefficient isgiven to different energy output types For an RIES theuncertainties of the source-side wind light and otheruncertain resources are large and have a great influence onthe stable operation of the system and the energy dispatchmanagement +erefore the source-side wind forecastingis set to a higher weight in addition as the power gridplays a leading role in the integrated energy system byvirtue of its perfect architecture and central hub advan-tages correspondingly the power load forecast is set to ahigher weight

+is paper proposes three calculation indicators namelymean absolute percentage error (MAPE) mean accuracy

x1 x2 xm

x1 x2 xn

x1 x2 xk

Hiddenlayer

Multitask learningOutputlayer

Inputlayer

Single-task learning

Wind powerhistorical data

Electrical loadhistorical data

Multisourcehistorical data

Figure 5 Comparison of single-task and multitasking learningbased on the LSTM network

6 Complexity

(MA) and weightedmean accuracy (WMA) as shown in thefollowing equation

MAPE 1n

1113944

n

i1

P(i) minus 1113954P(i)

P(i)

11138681113868111386811138681113868111386811138681113868

11138681113868111386811138681113868111386811138681113868times 100

MA 1 minus MAPE

WMA λ1MA1 + λ2MA2 + middot middot middot + λkMAk

(3)

In the equation P(i) and 1113954P(i) respectively representthe actual energy value and the predicted energy value at thefirst moment n is the sample data amount MAk representsthe prediction accuracy of the k prediction task and λk

represents the weight coefficient of the k prediction task Inthis paper the weight ratios of the energy predictions onboth sides of the source and the load are set as shown inTable 1

6 Simulation Analysis

61 Data Description +e experimental data in this paperare derived from the measured data provided by RIES in acertain area of Tianjin +e RIES consists of a power systema thermal system and a natural gas system +e source sideincludes energy supply equipment such as wind turbinesand photovoltaic power generation equipment Couplinglinks include energy conversion equipment such as cold andthermal power supply equipment electric boilers and gasboilers Energy demand mainly includes power load heatload and cooling load +e load side and the energy storageside contain devices such as electrical energy storage andthermal energy storage to improve the flexibility of thesystem In addition in order to prove the validity of thesource-load vertical information interaction the historicaldata of the lighting load reflecting the typical electricity

usage behavior of the user are also sampled All historicalenergy data sampling interval is 1 h data from July 2015 toJune 2016 are used as training set data and data from July2016 are used as test set data with 1 h as time step based onthe historical data of the first 7 days the energy supply andenergy demand on both sides of the source and the load inthe next day are predicted +e total 57168 samples weredivided into two parts (92 as the train set and 8 as the testset) +e forecasting models were trained by the train datasetand validated on the test dataset +e statistical informationof the above dataset is shown in Table 2

62 Analysis of Examples +is section analyzes the effect ofRIES multienergy information interactive energy prediction+e selected sunny day and rainy day of the test set wereanalyzed separately +e energy prediction results of windpower output photovoltaic output electric load heat loadand cold load on both sides of the source and the load in thetwo scenarios are shown in Figures 7 and 8 respectively

+e prediction results of Figures 7 and 8 show that theproposed source-load parallel LSTM learning model canobtain ideal results in wind energy photovoltaic electricload thermal load and cold load energy prediction in sunnyday scenario and rainy day scenario which confirms theeffectiveness of the model proposed in this paper +eprediction effect of each subtask in the parallel predictionresults shows a large difference Compared with the windand light prediction on the source side the predicted curvesof the charge heat and cold load on the charge side arecloser to the actual sequence curve which is due to theintermittent and random nature of source-side wind powerand photovoltaics

Comparing Figures 7(c)ndash7(e) it can be found that theelectric load is more volatile than the thermal load and thecold load which is because the thermal inertia and cold

LSTM 1

LSTMn

Forecast 1input data

Forecast ninput data

Parameters sharing

Forecast 1output data

Forecast noutput data

Inputssharing

Use

r beh

avio

r

Mul

tiene

rgy

sour

ce-lo

ad in

tegr

ated

fore

casti

ng re

sults

Figure 6 Interactive energy prediction model of RIES information based on parallel LSTM

Complexity 7

inertia of the thermal system are stronger [39] and thevolatility is also even smaller +ermal inertia means that thedynamic change of the thermal system is a slow process froma time scale Compared with the power system the heatenergy transmission exhibits a delay effect and the upperand lower fluctuations of the heat load exhibit an inertialeffect and the cold inertia is similar Because of the existenceof thermal inertia and cold inertia the thermal load in thethermal system is less random in the short-term predicted

time scale showing a characteristic of smooth variation soits prediction uncertainty is also small and the thermal loadand the cold load are relative to the electricity +e pre-diction error of the load is also relatively small ComparingFigures 7 and 8 it can be found that the uncertainty of actualphotovoltaic power and actual wind power in rainy dayscenes is larger than that of sunny days+e forecast curve ofphotovoltaic and wind power in sunny scenes is closer to thetrue value the electric load and heat load in rainy scenes

Table 1 Weight settings of source-load double-sided energy prediction tasks

Prediction task Wind power PV power Electric load Heat load Cold loadWeight coefficient 025 025 02 015 015

Table 2 +e description of simulation dataset

Data type Max (kW) Median (kW) Min (kW) Mean (kW) Std (kW) Train set size Test set sizeWind power 317561 233774 36612 216102 78964 8784 744Photovoltaic power 235000 0000 0000 42507 70557 8784 744Electric load 696585 584923 209872 523803 139174 8784 744Heat load 148473 92719 60532 100076 27240 8784 744Cold load 436964 223213 39046 221949 135304 8784 744Lighting load 03333 50000 2778 44792 27734 8784 744Multisource fusion mdash mdash mdash mdash mdash 52704 4464

True valuePredictive value

50

100

150

200

250

300

Win

d po

wer

(kW

)

5 10 15 20 250t (h)

(a)

True valuePredictive value

5 10 15 20 250t (h)

0

50

100

150

200

250

PV p

ower

(kW

)

(b)

True valuePredictive value

5 10 15 20 250t (h)

200

300

400

500

600

700

Elec

tric

load

(kW

)

(c)

True valuePredictive value

5 10 15 20 250t (h)

40

60

80

100

120

140

Hea

d lo

ad (k

W)

(d)

True valuePredictive value

5 10 15 20 250t (h)

0

100

200

300

400

500

Col

d lo

ad (k

W)

(e)

Figure 7 +e results of source-load bilateral energy prediction within a sunny day (a) Prediction of wind power (b) Prediction of PVpower (c) Prediction of electric load (d) Prediction of heat load (e) Prediction of cold load

8 Complexity

should increase as a whole +is is due to the cold weather inthe rainy scenes which affects the userrsquos energy use behaviorand increases the userrsquos demand for electricity and heat thusincreasing the electrical load and heat load

621 Comparative Analysis of Multienergy Load InformationInteraction In order to verify the effect of interactivelearning of horizontal and vertical multienergy informationthe parallel LSTM predictions considering user energy be-havior and parallel LSTM prediction without consideringuser energy behavior and the results of single LSTM pre-diction without considering user energy behavior are pro-posed to conduct a comparative analysis In order to ensurethe fairness of the experimental test this paper uses the samedataset for training and testing

+e comparison experiment setup is shown in Table 3 Inthe parallel LSTM learning without considering the userrsquosenergy behavior the input sample is based on the originalprediction model A and the lighting load data reflecting theuser behavior in the input sample are not considered +eprediction is based only on the historical data of the pre-dicted object the model C does not consider the singleLSTMprediction of the userrsquos energy behavior and the inputand output of each subprediction model only consider theinput and output of a single energy and there is also noparameter sharing between the LSTM unit of the different

submodels Model B uses parallel LSTM learning +ere isparameter sharing among each subtask which belongs to amultitask learning Model C uses a single LSTM learningand each subtask does not affect each other It belongs to asingle-task learning

+e prediction accuracy of the multienergy charge in-formation interaction comparison experiment is shown inTable 4 and Figure 9 +rough the analysis of Table 4 andFigure 9 it shows that the average prediction accuracy of theweights of the models A B and C are 09009 08917 and08728 respectively +e prediction result of the model A isthe best the model B is the second and the model C is thethird +at is considering the interaction of vertical andhorizontal information the energy predictions on both sidesof the source and the load show better prediction accuracy

Comparing the prediction results of Model B and ModelC for the five subprediction tasks when the different qualityenergies are predicted together in parallel a higher precisionthan the single prediction is obtained It can be seen thatparallel LSTM learning considering horizontal multienergyinformation interaction utilizes the coupled data betweendifferent systems to fully exploit the interactive informationbetween different quality energy sources of the integratedenergy system so that the prediction performance of acertain energy can be improved In addition considering themultienergy horizontal information interaction the pre-dicted accuracy of the load-side electrical load heat load and

True valuePredictive value

5 10 15 20 250t (h)

100

150

200

250

300

350W

ind

pow

er (k

W)

(a)

True valuePredictive value

5 10 15 20 250t (h)

0

20

40

60

80

100

120

PV p

ower

(kW

)(b)

True valuePredictive value

5 10 15 20 250t (h)

300

400

500

600

700

Elec

tric

load

(kW

)

(c)

True valuePredictive value

5 10 15 20 250t (h)

60

80

100

120

140

160

Hea

t loa

d (k

W)

(d)

True valuePredictive value

5 10 15 20 250t (h)

0

100

200

300

400

500

Col

d lo

ad (k

W)

(e)

Figure 8+e results of source-load bilateral energy prediction within a rainy day (a) Prediction of wind power (b) Prediction of PV power(c) Prediction of electric load (d) Prediction of heat load (e) Prediction of cold load

Complexity 9

cold load are increased by 313 253 and 267 re-spectively while the predicted accuracy of the source-sidewind power and photovoltaic is increased by 152 and146 and the energy prediction accuracy of the charge sideis relatively high It can be seen that the source-side un-certainty resources are also affected by other energy in-teractions in the system but the impact is small while thecharge side demand is largely affected by other energy in-teractions in the system And the information interaction ismore intimate between each other

Comparing the prediction results of Model A and ModelB when the userrsquos energy behavior information interactsbetween the source and the load in vertical direction theaverage weight accuracy of the overall prediction is increasedby 103 Comparing the performance improvement effectsof the five subprediction tasks it can be found that thepercentage improvement of PV prediction accuracy is185 and the improvement effect is relatively best fol-lowed by the electrical load while the prediction accuracy ofwind power heat load and cooling load is not obvious Itcan be seen that considering the source-load vertical in-formation interaction can improve the energy predictionaccuracy on both sides of the source and the load but thespecific user energy behavior has different differences in theenergy prediction of different types of users and the user

lighting behavior is obviously related to photovoltaic powergeneration and user electricity +e interaction betweenloads is more tight and its effect on PV prediction andelectric load forecasting is more obvious

622 Comparison of Deep Neural Network Algorithms withOther Algorithms In order to verify the prediction effect ofthe LSTM deep neural network algorithm the LSTM modeland the autoregressive integrated moving average model(ARIMA) algorithm and the classic BP neural network(BPNN) algorithm in this paper are proposed +e simu-lation results of them were compared and analyzed In orderto ensure the fairness of the experimental test this paper usesthe same dataset for training and testing and both use asingle-task prediction strategy +e prediction results of thethree algorithms are shown in Table 5 and Figure 10

Observing the prediction results the prediction effect ofARIMA algorithm is the most general and the predictionaccuracy of BPNN algorithm is much higher than that ofARIMA algorithm +e prediction accuracy of each task ofLSTM deep learning prediction algorithm proposed in thispaper is the highest +e ARIMA algorithm can not effec-tively use the historical information in the long-term se-quence because it only uses a part of the historicalinformation to train which leads to the large predictionerror +e LSTM algorithm can fully utilize its strongnonlinear learning ability and long-term and short-termmemory advantages to exploit the information value of long-term historical data so the grasp of the energy prediction lawis more precise

In addition comparing the energy prediction results ofthe LSTM algorithm and the BPNN algorithm one by onethe LSTM algorithm predicts that the accuracy of thethermal load and the cold load prediction is more obvious Itcan be seen that due to the greater inertia of the thermalsystem the time scale of the change law is longer whichmakes the prediction accuracy of the long-term and short-term deep learning on the heat load and the cold loadsignificantly higher than the electric load

In order to evaluate the computational efficiency ofalgorithms four forecasting scenarios are performed tentimes so that the forecasting results can reliably be sum-marized as in Table 6 Parallel LSTM and Single LSTM havelonger training time but the prediction time is kept within

Table 3 Contrast experiment setting of multisource source information interaction

Model Vertical interaction Horizontal interactionModel A User energy behavior is considered Parallel LSTM learningModel B User energy behavior is not considered Parallel LSTM learningModel C User energy behavior is not considered Single-LSTM learning

Table 4 +ree kinds of model prediction results of information interaction comparison experiment

Precision Wind power PV power Electric load Heat load Cold load Mean precisionModel A 08602 08735 09236 09445 09407 09009Model B 08542 08576 09113 09401 09368 08917Model C 08414 08453 08836 09169 09124 08728

Model AModel BModel C

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

PV Electric Heat Cold MeanWindPrediction task

Figure 9 Comparison of prediction effects of three models

10 Complexity

one minute and can meet the requirements of short-termforecasting Further analysis shows that the parallel trainingtime and prediction time of parallel LSTM is greatly im-proved compared to single LSTM due to the combination offive subtasks into one multitask learning In summary al-though the LSTM algorithm is time consuming in trainingtime for short-term prediction the training process doesnot require online completion and the prediction time ofthe LSTM algorithm especially parallel LSTM can fullymeet the requirements of short-term prediction Consid-ering the prediction accuracy the sacrifice of the LSTMalgorithm in computing time is worthwhile

7 Conclusion

In this paper the horizontal and vertical information interactioncharacteristics between multiple energy loads in RIES areconsidered Using LSTM network deep learning algorithm a

parallel LSTM-based multienergy-load information interactiveshort-term energy prediction method is proposed+rough thesimulation of the test data the following conclusions are drawn

(1) +e ldquoinformation interactionrdquo strategy proposed inthis paper can effectively improve the predictionperformance of source-load energy prediction inregional integrated energy systems

(2) Parallel prediction considering horizontal in-teraction mainly affects the prediction effect on theload side while the information sharing of user il-lumination behavior considering vertical interactionmainly affects the improvement of PV predictioneffect

(3) Compared with the conventional prediction algo-rithm the LSTM deep learning algorithm can learnthe deep hidden information of historical dataimprove the prediction accuracy while meetingcalculation time and prove the effectiveness of thealgorithm in the prediction of time series featuresequence

However the performance of the proposed parallelLSTM algorithm may be further enhanced by some evo-lutionary algorithms or multiobjective optimization algo-rithms such as differential evolution algorithm [40]multiobjective extremal optimization [41] and evolutionarymultiobjective optimization algorithms [42ndash47] Of coursesome significant topics are worth researching on the opti-mization of the proposed parallel LSTM by multiobjectiveoptimization

Data Availability

+e data used to support the findings of this study areavailable from the corresponding author upon request

Conflicts of Interest

+e authors declare no potential conflicts of interest

Acknowledgments

+is work was fully supported by the National Key Researchand Development Program of China (no 2018YFB0904200)

References

[1] A Q Huang M L Crow G T Heydt J P Zheng andS J Dale ldquo+e future renewable electric energy delivery andmanagement (FREEDM) system the energy internetrdquo Pro-ceedings of the IEEE vol 99 no 1 pp 133ndash148 2011

[2] I Colak S Sagiroglu G Fulli M Yesilbudak and C CovrigldquoA survey on the critical issues in smart grid technologiesrdquo

Table 5 Comparison of prediction results of three algorithms

Precision Wind power PV power Electric load Heat load Cold load Mean precisionLSTM 08414 08453 08836 09169 09124 08728BPNN 08216 08314 08695 08912 08889 08542ARMIA 07733 07824 08282 08454 08408 08075

Prediction precision ofwind powerPrediction precision ofPV powerPrediction precision ofelectric load

Prediction precision ofhead loadPrediction precision ofcold loadWeight average predictionprecision

07

075

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

BPNN ARMIALSTMPrediction algorithm

Figure 10 Comparison of prediction effects of three algorithms

Table 6 Comparison of computational efficiency of fouralgorithms

Algorithm Training timeaverage (sec)

Prediction timeaverage (sec)

Parallel LSTM 436374 1002Single LSTM 1218337 3662BPNN 1094911 2253ARMIA 54696 215

Complexity 11

Renewable and Sustainable Energy Reviews vol 54 pp 396ndash405 2016

[3] C Tu X He Z Shuai and F Jiang ldquoBig data issues in smartgrid-a reviewrdquo Renewable and Sustainable Energy Reviewsvol 79 pp 1099ndash1107 2017

[4] Y Zhang R Yang K Zhang H Jiang and J J ZhangldquoConsumption behavior analytics-aided energy forecastingand dispatchrdquo IEEE Intelligent Systems vol 32 no 4pp 59ndash63 2017

[5] J Deng F Wang Y Chen and X Zhao ldquoFrom industry 40 toenergy 50 the concept connotation and system framework ofintelligent energy systemsrdquo Journal of Automation vol 41no 12 pp 2003ndash2016 2015

[6] Q Sun L Shi Y Ni D Si and J Zhu ldquoAn enhanced cas-cading failure model integrating data mining techniquerdquoProtection and Control of Modern Power Systems vol 2 2017

[7] D Zheng A T Eseye J Zhang and H Li ldquoShort-term windpower forecasting using a double-stage hierarchical ANFISapproach for energy management in microgridsrdquo Protectionand Control of Modern Power Systems vol 2 pp 1ndash10 2017

[8] Z Li L Ye Y Zhao X Song J Teng and J Jin ldquoShort-termwind power prediction based on extreme learning machinewith error correctionrdquo Protection and Control of ModernPower Systems vol 1 no 1 pp 1ndash8 2016

[9] H Z Wang Z X Lei X Zhang B Zhou and J C Peng ldquoAreview of deep learning for renewable energy forecastingrdquoEnergy Conversion and Management vol 198 pp 1ndash16 2019

[10] H Z Wang G B Wang G Q Li J C Peng and Y T LiuldquoDeep belief network based deterministic and probabilisticwind speed forecasting approachrdquo Applied Energy vol 182pp 80ndash93 2016

[11] M-R Chen G-Q Zeng K-D Lu and J Weng ldquoA two-layernonlinear combination method for short-term wind speedprediction based on ELM ENN and LSTMrdquo IEEE Internet ofings Journal vol 6 no 4 pp 6997ndash7010 2019

[12] H Liu X Mi and Y Li ldquoSmart multi-step deep learningmodel for wind speed forecasting based on variational modedecomposition singular spectrum analysis LSTM networkand ELMrdquo Energy Conversion and Management vol 159pp 54ndash64 2018

[13] J Chen G-Q Zeng W Zhou W Du and K-D Lu ldquoWindspeed forecasting using nonlinear-learning ensemble of deeplearning time series prediction and extremal optimizationrdquoEnergy Conversion and Management vol 165 pp 681ndash6952018

[14] W Kong Z Y Dong Y Jia D J Hill Y Xu and Y ZhangldquoShort-term residential load forecasting based on LSTM re-current neural networkrdquo IEEE Transactions on Smart Gridvol 10 no 1 2017

[15] H Gao W Huang X Yang Y Duan and Y Yin ldquoTowardservice selection for workflow reconfiguration an interface-based computing solutionrdquo Future Generation ComputerSystems vol 87 pp 298ndash311 2018

[16] L Qi Q He F Chen et al ldquoFinding all you need web APIsrecommendation in web of things through keywords searchrdquoIEEE Transactions on Computational Social Systems vol 6no 5 2019

[17] L Qi R Wang C Hu S Li Q He and X Xu ldquoTime-awaredistributed service recommendation with privacy-preserva-tionrdquo Information Sciences vol 480 pp 354ndash364 2019

[18] Y Zhang and Q Yang ldquoAn overview of multi-task learningrdquoNational Science Review vol 5 no 1 pp 34ndash47 2018

[19] A Argyriou M T Evgeniou and M Pontil ldquoConvex multi-task feature learningrdquo Machine Learning vol 73 no 3pp 243ndash272 2008

[20] T Jebara ldquoMultitask sparsity via maximum entropy dis-criminationrdquo Journal of Machine Learning Research vol 12no 1 pp 75ndash110 2011

[21] L Qi Y Chen Y Yuan S Fu X Zhang and X Xu ldquoA QoS-aware virtual machine scheduling method for energy con-servation in cloud-based cyber-physical systemsrdquo WorldWide Web vol 5 pp 1ndash23 2019

[22] R Zhang P Xie C Wang G Liu and S Wan ldquoClassifyingtransportation mode and speed from trajectory data via deepmulti-scale learningrdquo Computer Networks vol 162 p 1068612019

[23] W Wang D Wang H Jia et al ldquoReview of steady-stateanalysis of typical regional integrated energy system under thebackground of energy internetrdquo Proceedings of the CSEE12pp 3292ndash3306 2016

[24] S Ding S Qu Y Xi and S Wan ldquoStimulus-driven andconcept-driven analysis for image caption generationrdquoNeurocomputing 2019

[25] P Vrba V Marik P Siano et al ldquoA review of agent andservice-oriented concepts applied to intelligent energy sys-temsrdquo IEEE Transactions on Industrial Informatics vol 10no 3 pp 1890ndash1903 2014

[26] W Jia C Kang and Q Chen ldquoAnalysis on demand-sideinteractive response capability for power system dispatch in asmart grid frameworkrdquo Electric Power Systems Researchvol 90 pp 11ndash17 2012

[27] Z C Lipton J Berkowitz and C Elkan ldquoA critical review ofrecurrent neural networks for sequence learningrdquo CoRR2015 httpsarxivorgpdf150600019pdf

[28] T N Sainath O Vinyals A Senior and H Sak ldquoCon-volutional long short-term memory fully connected deepneural networksrdquo in Proceedings of the 2015 IEEE In-ternational Conference on Acoustics Speech and Signal Pro-cessing (ICASSP) pp 4580ndash4584 IEEE Brisbane AustraliaApril 2015

[29] Y Guo Z Cheng L Nie YWang J Ma andM KankanhallildquoAttentive long short-term preference modeling for person-alized product searchrdquo ACM Transactions on InformationSystems (TOIS) vol 37 no 2 p 19 2019

[30] A Graves ldquoLong short-term memoryrdquo Neural Computationvol 9 pp 1735ndash1780 1997

[31] F A Gers J Schmidhuber and F Cummins ldquoLearning toforget continual prediction with LSTMrdquo Neural Computa-tion vol 12 no 10 pp 2451ndash2471 2000

[32] M L Seltzer and J Droppo ldquoMulti-task learning in deepneural networks for improved phoneme recognitionrdquo inProceedings of the 2013 IEEE International Conference onAcoustics Speech and Signal Processing pp 6965ndash6969 IEEEVancouver Canada May 2013

[33] S Wan Y Zhao T Wang Z Gu Q H Abbasi andK-K R Choo ldquoMulti-dimensional data indexing and rangequery processing via Voronoi diagram for internet of thingsrdquoFuture Generation Computer Systems vol 91 pp 382ndash3912019

[34] Z Gao D Y Wang S H Wan H Zhang and Y L WangldquoCognitive-inspired class-statistic matching with triple-con-strain for camera free 3D object retrievalrdquo Future GenerationComputer Systems vol 94 pp 641ndash653 2019

[35] C A Kang A R Brandt and L J Durlofsky ldquoOptimaloperation of an integrated energy system including fossil fuel

12 Complexity

power generation CO2 capture and windrdquo Energy vol 36no 12 pp 6806ndash6820 2011

[36] Q Zhang SWan BWang DW Gao and HMa ldquoAnomalydetection based on randommatrix theory for industrial powersystemsrdquo Journal of Systems Architecture vol 95 pp 67ndash742019

[37] Z Gao H Z Xuan H Zhang S Wan and K K R ChooldquoAdaptive fusion and category-level dictionary learningmodel for multi-view human action recognitionrdquo IEEE In-ternet of ings Journal p 1 2019

[38] WWu K Chen Y Qiao and Z Lu ldquoProbabilistic short-termwind power forecasting based on deep neural networksrdquo inProceedings of the 2016 International Conference on Proba-bilistic Methods Applied to Power Systems (PMAPS) pp 1ndash8Beijing China October 2016

[39] K M Powell A Sriprasad W J Cole and T F EdgarldquoHeating cooling and electrical load forecasting for a large-scale district energy systemrdquo Energy vol 74 pp 877ndash8852014

[40] A Peimankar S J Weddell T Jalal and A C LapthornldquoMulti-objective ensemble forecasting with an application topower transformersrdquo Applied Soft Computing vol 68pp 233ndash248 2018

[41] G-Q Zeng J Chen L-M Li et al ldquoAn improved multi-objective population-based extremal optimization algorithmwith polynomial mutationrdquo Information Sciences vol 330pp 49ndash73 2016

[42] Y-L Hu and L Chen ldquoA nonlinear hybrid wind speedforecasting model using LSTM network hysteretic ELM andDifferential Evolution algorithmrdquo Energy Conversion andManagement vol 173 pp 123ndash142 2018

[43] G-Q Zeng J Chen Y-X Dai L-M Li C-W Zheng andM-R Chen ldquoAn improved Design of fractional order PIDcontroller for automatic regulator voltage system based onmulti-objective extremal optimizationrdquo Neurocomputingvol 160 pp 173ndash184 2015

[44] X Xu R Gu F Dai L Qi and S Wan ldquoMulti-objectivecomputation offloading for internet of vehicles in cloud-edgecomputingrdquo Wireless Networks pp 1ndash19 2019

[45] S Wan X Li Y Xue W Lin and X Xu ldquoEfficient com-putation offloading for internet of vehicles in edge comput-ing-assisted 5G networksrdquo e Journal of Supercomputingpp 1ndash30 2019

[46] B Wang S Wan X Zhang and K-K R Choo ldquoA novelindex for assessing the robustness of integrated electricalnetwork and a natural gas networkrdquo IEEE Access vol 6pp 40400ndash40410 2018

[47] Y Yin F Yu Y Xu L Yu and J Mu ldquoNetwork location-aware service recommendation with random walk in cyber-physical systemsrdquo Sensors vol 17 no 9 p 2059 2017

Complexity 13

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Page 2: Parallel LSTM-Based Regional Integrated Energy System ...downloads.hindawi.com/journals/complexity/2019/7414318.pdf · Characteristics of Regional Integrated Energy System “Information

data analysis and prediction algorithms to verify the cor-relation between energy consumption behavior and dis-tributed PV output and input the prediction results into thepredicted energy management system to optimize load-sideenergy consumption and balance the systemrsquos energy supplyand load demand Deng et al [5] proposed the energy 50concept based on cyber-physical-social systems (CPSS) andanalyzed the social behavior characteristics of load-sideenergy consumption in smart home energy systems bycollecting a variety of operational information in the actualsystem by using the ldquodata driven + physical modelrdquo approachto form a new energy forecasting agent to achieve optimaloperational control [6ndash8] In addition in the energy pre-diction of RIES deep learning prediction technology hasattracted the attention of many scholars at home and abroadWang et al [9] have made extensive analysis and discussionon renewable energy forecasting methods based on deeplearning discussed the current problems facing this di-rection and prospected the future development directionWang et al [10] proposed a wind speed prediction modelbased on deep confidence network and achieved goodprediction results Chen et al [11] proposed a new two-layernonlinear combination short-term wind speed predictionmethod (EEL-ELM) In the first layer the extreme learningmachine (ELM) Elman neural network (ENN) and longshort-term memory (LSTM) neural network are used topredict wind speed and three prediction results are ob-tained +en the nonlinear aggregation mechanism of ex-treme learning machine (ELM) is used to alleviate theinherent weakness of single method and linear combinationLiu et al [12] combined variational mode decomposition(VMD) singular spectrum analysis (SSA) long short-termmemory (LSTM) network and extreme learning machine(LEM) and proposed a new multistep wind speed predictionmodel +e simulation results show that the method is moreeffective and robust in extracting trend information On thebasis of multiple long short-term memory (LSTM) learningthe literature [13] combines the extreme value optimizationalgorithm and the support vector machine model to in-tegrate the LSTM layer prediction results and predicts thewind speed in a short time In literature [14] the LSTM deepneural network algorithm is used and the userrsquos energy dataof various types of equipment are used to predict the res-idential load on the load side in a short term which provesthe data value of the load energy information in the fore-casting field Multitask learning has been studied for about20 years [15ndash17] At present multitask learning methods canbe roughly summarized into two categories One is to sharethe same parameters between different tasks and the other isto mine the shared data features hidden between differenttasks [18] Argyriou et al [19] detailed the four multitaskinglearning methods of feature selection kernel selectionadaptive pooling and graphical model structure Jebara [20]proposed a typical multitask model for mining commonfeatures between multitasks A framework for multitaskfeature learning is given in this paper which becomes thebasis for many multitask learning references [21 22] In-spired by the above literature this paper argues that it isnecessary to explore the value of multienergy coupling

information and load-side user behavior characteristics forenergy system operation and use the information correlationcharacteristics between multiple energy sources to expandthe concept of multienergy interaction to the forecastingstage +e ldquoinformation interactionrdquo strategy is proposedand explored the data information implied in each link ofRIES by combining deep learning technology and the en-ergy utilization efficiency is maximized through multienergyinteraction Based on the existing research this papercomprehensively considers the horizontal and vertical in-formation interaction characteristics between multiple en-ergy loads in RIES and proposes a parallel energy LSTMbased on multienergy-load information interactive short-term energy prediction strategy Firstly the RIES multi-energy-charge coupling correlation is analyzed and thehorizontal and vertical ldquoinformation interactionrdquo modes ofmultienergy load are proposed +en from the perspectiveof information interaction based on the LSTM deep neuralnetwork time series prediction the parallel LSTM multitasklearning model is established to realize the horizontal in-teraction between multienergy systems At the same timebased on the userrsquos energy behavior data driving the verticalinteraction between the source and the load is realized andthe uncertain sources are used for integrated energy pre-diction of the wind power photovoltaic and various loadson both sides of the source and the load Finally based on themeasured data of a regionrsquos RIES the prediction model istrained and verified +e simulation results show that theproposed information interactive parallel prediction methodcan effectively improve the prediction effect of each subtask

2 Regional Integrated Energy System

21 Basic Structure of Regional Integrated Energy System+e regional integrated energy system takes the electricity asthe core is based on smart grid and is led by clean energy Ituses advanced information communication and energyconversion technologies to organically connect the links ofproducing transporting storing and consuming in themultienergy systems such as electricity gas and heat so as toachieve optimal allocation of energy multienergy couplingand complement [23] RIES structurally includes regionalenergy production input network multienergy couplingnetwork and regional internal load-side energy outputnetwork It can fully absorb all kinds of renewable energysuch as wind power and photovoltaic and access varioustypes of loads such as electricity gas heat and cold toachieve efficient energy allocation At the same time its toplayer has an integrated information communication supportsystem to achieve the interactive fusion of energy flow andinformation flow +e basic architecture of the RIES de-scribed in this paper is shown in Figure 1

22 Characteristics of Regional Integrated Energy SystemldquoInformation Interconnectionrdquo Under the support of theenergy Internet and multienergy key technologies charac-terized by ldquoopenrdquo and ldquointerconnectedrdquo the coupling in-teraction of multienergy-network-load-storage in RIES is

2 Complexity

more tight [24] (1) From the perspective of horizontalinterconnection RIES will integrate the energy subsystemssuch as coldheatelectricgastraffic to construct a mul-tienergy flow system +e energy is converted and flowedbetween the systems through the coupling componentsand the various energy forms complement each other +eenergy resources have to be optimally configured in a widearea that is to achieve the ldquohorizontal information in-terconnectionrdquo of the energy Internet (2) From the per-spective of vertical interconnection based on thetraditional energy supply network RIES will realize thereal-time sharing of information resources in the source-network-load-storage through Internet technology +isresource sharing mechanism has two-way conductioncharacteristics +e user has information bidirectionalconduction capability between the information controlsystem the energy supply module and the multienergyintelligent delivery module +e interactive sharing ofenergy information can also be realized between users thatis to realize the ldquovertical information interconnectionrdquo ofthe energy Internet

3 RIES Multienergy CouplingCorrelation Analysis

31 RIES Horizontal Multienergy Information Coupling+e horizontal multienergy information coupling corre-lation of RIES is represented by the correlation betweendifferent energy flow subsystems such as electricity gasand heat

From the perspective of coupling the composition of theRIES can be divided into a coupling unit and a noncouplingunit As shown in Figure 2 the coupling unit connectsdifferent energy flow subsystems and consumes one or moreenergy sources to generate other energy sources Typicalcoupling forms are as follows cold and heat electricity

supply units consume coal or natural gas to generate elec-tricity heat and cold +e heat pump consumes electricalenergy to generate thermal energy the electrical hydrogenconsumes electrical energy to generate hydrogen and thecoupling unit causes the energy flow subsystems to interact+e uncoupled unit includes the network and equipmentinside each energy flow subsystem and other uncoupledboundary conditions such as source and load +e uncou-pled unit also affects other systems through the coupling unit[25]+erefore different energy flow subsystems have strongcoupling between them and there is a large amount ofhorizontal multienergy coupling information existing in themultienergy system data resulting in great correlation be-tween various units within the system

32 RIES Vertical Source-Load Information Coupling +evertical source-load information coupling correlation ofRIES is represented by the correlation between the multi-energy supply of the source and the userrsquos energy behavior of

WT

Input side

Distributionnetwork

Natural gas WindSolar

Output side Cold load Electric

loadHeat load

PV

Refrigeration

CCHPGB

ES

HS

Coupling side

Informationcommunicationsupport system

Calculationdata storage

layer

Informationcorrespondence

network

Acquisitionmonitoring

layer

Analysisdecision level

Lateral information interaction between multiplecapabilities

Vert

ical

info

rmat

ion

inte

ract

ion

betw

een

sour

ce lo

ads

sum

sum

Figure 1 Basic framework of RIES

Gas

Gasturbine

CHP

Powerto gas

Electric storage

Heatpump

Gasboiler

HeatCirculating pump

Heat storage

Gas storage

Electric

Figure 2 Transverse multicoupling schematic diagram

Complexity 3

the load Different energy behaviors correspond to differentload characteristics [26] For example the lighting behaviorin RIES is a typical user energy behavior +rough advancedmeasurement devices the userrsquos lighting behavior can bereflected as a lighting load curve

From the perspective of user energy behavior there aremultiple coupling relationships between multiple energyloads in RIES Taking the userrsquos lighting behavior as anexample if the cloud suddenly covers a certain area andreduces the ambient light intensity it will affect the userrsquoslighting behavior people will increase the lighting equip-ment usage to meet the lighting demand and the lightingload will increase At the same time such cloud movementswill also have an impact on photovoltaic power generationcloud clusters will block the photovoltaic panels which willaffect the irradiance reduction on the photovoltaic panelsand affect the photovoltaic power generation In order toprove the characteristics of this behavioral feature this papertakes the data of three days from January 24 to 26 2016 ofRIES in a certain area of Tianjin for quantitative analysis+ephotovoltaic power output curve and the user lighting loadcurve are shown in Figure 3

+e upper part of Figure 3 is the curve trend of thephotovoltaic power output and the lower part is the curvetrend of the userrsquos lighting load (1) Both the userrsquos lightingload and the photovoltaic power output have a day and nightcycle +e distribution of them is mainly concentrated in thedaytime +e peak of the userrsquos lighting load is generallydistributed between 18 00 and 20 00 and the peak outputof the photovoltaic power supply is generally distributedbetween 12 00 and 14 00 at noon (2) Taking the three-dayshort-term data for comparison analysis the overall outputof photovoltaic power generation on the third day wassignificantly larger than the previous two days Accordinglythe lighting load on the third day between 10 00 and 16 00is reduced compared to the previous two days and it isreduced to the low of the day It shows that when the outputof photovoltaic power generation increases it also affects theuserrsquos energy consumption behavior when the ambient lightintensity is large which means that the userrsquos demand forlighting energy is reduced (3) Take the ultra-short-termdata analysis from 12 00 to 16 00 on the first day andthe photovoltaic power output reaches the peak of one day at13 00 and the lighting load is also reduced to low of thedaytime at this moment During the period from 13 00 to14 00 a sudden downward climb occurs and the lightingload increases at the corresponding time +is ldquomutationrdquophenomenon should be caused by the cloudy weather af-fected by the cloud From the above analysis it can be foundthat there is a strong information coupling correlation be-tween the ldquosourcerdquo photovoltaic power output and theldquocharge endrdquo userrsquos lighting load

33 RIES Multienergy Information Interaction ModeRIES has both horizontal multienergy information couplingcorrelation characteristics and vertical source-load in-formation coupling correlation characteristics From theperspective of information interaction energy management

there are also two horizontal and vertical information in-teraction modes

+e horizontal information interaction mode is based onmultienergy conversion information sharing On the basis ofobtaining the measurement data of the multienergy flowsubsystem the nonlinear artificial intelligence deep learningtechnology is used to process the multienergy data in par-allel and it effectively utilizes the complex shared in-formation of energy conversion identifies the abstractfeatures of the training data and improves the precision ofsubtask prediction of the energy management

+e vertical information interaction mode is based onuser behavior information sharing On the basis of obtainingmulticategory data on both sides of the source and the loadthe deep learning method is used to dig the user-side be-havior characteristics and the associated user behaviorinformation data are taken as input to guide the source-sideenergy management prediction and improve the predictionaccuracy of the energy supply end

4 Multitask Learning Mechanism for DeepLearning Energy Prediction

41 LSTM Recurrent Neural Network

411 Principle of LSTM Cyclic Neural Network +e trainingobjects of the source-charged uncertain resources haveobvious timing characteristics and the sequence data havestrong correlation before and after From the perspective oflearning and training recurrent neural network (RNN) canprocess sequences with temporal correlation of arbitrarylength by using neurons with self-feedback which has ob-vious advantages in processing sequence data [27] +estructure of the RNN is shown in Figure 4(a) +e LSTMnetwork is a variant of RNN that overcomes the traditionalRNN gradient demise problem [28] +e hidden layer is nolonger a simple neural unit but an LSTM unit with a uniquememory mode Each LSTM unit contains a state cell thatdescribes the current state of the LSTM unit [29] c repre-sents the current state quantity of the LSTM unit and h

represents the current output of the LSTM unitLSTM reads and modifies state units by controlling the

forget gates input gates and output gates [30] which are

0

ndash100

250

50

100

150

200

ndash50

ndash200

ndash100

0

100

200

300

P (k

W)

10 20 30 40 50 60 700t (h)

Figure 3 +e relationship between the photovoltaic power curveand the lighting load curve of the user

4 Complexity

typically formed by sigmoid or tanh functions and Hada-mard product operation As shown in Figure 4(b) the forgetgate determines ctminus 1 of the previous moment and how manycomponents of the cell state ct remain at the current timeand the input gate determines how many components of thenetwork input are saved to the cell state ct at the current timext and howmany components of the output gate control cellstate ct are output to the current value ht of the LSTM

+e calculation formula between the variables of theLSTM unit is as follows

ft σ Wfxxi + Wfhhtminus 1 + bf1113872 1113873

it σ Wixxi + Wihhtminus 1 + bi( 1113857

gt ϕ Wgxxt + Wghhtminus 1 + bg1113872 1113873

ct ft ⊙ ctminus 1 + it ⊙gt

ot σ Woxxt + Wohhtminus 1 + bo( 1113857

ht ot ⊙ϕ ct( 1113857

(1)

In the formulaWfxWfhWixWihWgxWghWox andWoh are the weight matrices corresponding to the input of thenetwork activation function bf bi bg and bo are the offsetvectors ⊙ represents the multiplication of the Hadamardproduct matrix σ represents the sigmoid activation functionand ϕ represents the tanh activation function

412 LSTMDeep Learning Training Algorithm +e trainingof LSTM network adopts BP-based backpropagationthrough time (BPTT) [31] +e algorithm steps are as fol-lows (1) Calculate the output value of each neuron forwardFor the LSTMunit the values of the six vectors such asft itgt ct ot and ht are calculated (2) Calculate the error termof each neuron in reverse Similar to traditional RNN thebackpropagation of the LSTM error term consists of twolevels one is at the spatial level and the error term ispropagated to the upper layer of the network the other is atthe time level which propagates back in time For examplefrom the current time of t the error at each moment iscalculated (3) Calculate the gradient of each weightaccording to the corresponding error term and update thenetwork weight parameter (4) Jump to the first step andcontinue to perform the first 3 steps until the network erroris less than the given value

42 LSTM Network Multitask Learning MechanismMultitask learning (MTL) as an inductive migrationmechanism avoids the underfitting of model training byusing the correlation between multiple tasks digging thehidden common features and improving the generalizationlearning ability of the model [32] RIES source wind andlight prediction and load-end electricity gas heat and coldload prediction can be regarded as different tasks in mul-titask learning [33] Multiple tasks share some common dataor model structures that is there is a certain correlationbetween the multienergy prediction tasks +e multitasklearning prediction model based on the LSTM networkproposed in this paper shares the hidden layer nodes amongall prediction tasks through the hard parameter sharingmethod and trains the training data corresponding to themultienergy prediction tasks at the same time consideringthe correlation information between tasks to improve thelearning ability of all predictive tasks [34] A comparison ofsingle-task and multitask learning based on the LSTMnetwork is shown in Figure 5 +e multisource historicaldata in the figure is a collection of five types of historical dataincluding wind power photovoltaic electric load heat loadand cold load

5 Parallel LSTM-Based RIES InformationInteractive Energy Prediction Method

51 RIES Multienergy Prediction Model Design RIES hasuncertainties at both the source and the load end so accurateprediction of uncertain resources is the basis for imple-menting RIES energy optimization management andscheduling [35] +e source mainly includes wind powerprediction and photovoltaic power prediction and the loadmainly includes electric load prediction thermal load pre-diction and cold load prediction

As described in the second section of this paper there areintricate mutual coupling relationships between multipleenergy loads of RIES and the prediction tasks of uncertainresources on both sides of the source and load have strongcorrelation It can effectively improve the overall predictionaccuracy of the RIES uncertain resources by learning thiscoupling relationship through joint prediction training [36]From the perspective of RIES multienergy-load informationinteraction this paper proposes the concept of ldquoparallel

xtndash1 xt xt+1

htndash2

ctndash2

hh h

cc c

htndash1 ht ht+1

ctndash1 ct ct+1

(a)

tanh

tanh

xt

ht

ctctndash1

Forgetgate

ft it ot

ct

htndash1

Inputgate

Outputgate

σ σ σ

(b)

Figure 4 Schematic diagram of RNN and LSTM (a) +e structure of the RNN (b) +e structure of the LSTM cell

Complexity 5

predictionrdquo and utilizes the strong nonlinear learning abilityand historical memory advantages of LSTM deep learningnetwork to conduct integrated prediction of uncertain re-sources at both source and load As shown in Figure 6 thehorizontal information interaction based on multienergyconversion information sharingmainly utilizes themultitasklearning parameter sharing mechanism of the parallel LSTMprediction network the vertical information interactionbased on the user behavior information sharing mainly usesthe user behavior data to guide source-load multitask pre-diction as an additional unified input information Para-metric sharing is weight sharing which takes advantage ofthe correlation between subtasks Input sharing refers todata sharing Different subtasks share historical power datasuch as wind power photovoltaic power power loadthermal load and cooling load

52 Data Preprocessing +e training samples select the windpower generation time series on the source side the photo-voltaic power generation time series the charge side on the loadside the heat load the cold load time series and the illumi-nation load series data Considering the input and output rangeof the nonlinear activation function in the model in order toavoid neuron saturation the data are normalized and mappedbetween [minus 1 1] as shown in the following equation

xprime x minus xmax + xmin( 11138572

xmax minus xmin( 11138572 (2)

In the equation x is the actual input or output data andxmax and xmin are the maximum and minimum values of thevariable respectively

53 Task Construction +e essence of using the multitasklearning method for multienergy-load integration predictionlies in the close correlation of each prediction subtask [37]+e traditional single-task prediction method uses a singledimension of historical observations to predict its futureoutput and the model output is a single attribute +emultitask learning integration prediction uses the multidi-mensional data on both sides of the source and the load as a

common input to effectively utilize the sample On the otherhand through the shared hidden layer node divisionmethodthe five attributes of wind photovoltaic power generationelectricity heat and cold load are simultaneously predictedto play the role of inductive bias and improve predictionaccuracy

54 Model Training +e LSTM-based source-charge in-tegrated prediction model mainly needs to determine thefollowing parameters input layer dimension input layertime step hidden layer number each hidden layer di-mension and output layer dimension [38]

+e time step of the input layer is generally the length ofthe variable time series used for the source-load integrationprediction According to the periodic characteristics of theprediction object sequence the input layer time step is set to24 that is the historical data whose resolution ratio is 1 hduring the first seven days before the input is used forforecasting and it takes up one input node per day+e inputlayer dimension is the number of nodes +e input layernode of this paper contains six objects wind power gen-eration time series photovoltaic power generation timeseries charge side load heat load cold load time series andlighting load series which constitutes totally 42 nodes so theinput layer dimension is set to 42 Since this paper simul-taneously predicts six attributes of wind power photovol-taic electricity heat and cold load the output layer has 5nodes that is the output layer dimension is set to 5 In thispaper based on experience and multiple trial and com-parison the number of hidden layers is set to 2 and thedimension of the hidden layer is set to 30 +e selection ofthe activation function is as shown in Figure 4(a) thebatch_size is set to 42 the numbers of iterations of an epochis set to 8784 the learning rate is set to 080 and the learningdelay rate is set to 005 +e network training function usesthe BPTT algorithm mentioned above

55 Evaluation Indicators Multienergy-load integrationforecasting simultaneously trains multiple forecastingtasks To comprehensively measure the forecasting effect itis necessary to evaluate the integrated forecasting accuracyas a whole [39] In this paper the average accuracyevaluation index based on weight coefficient is proposedFor the importance degree of different energy on the twosides of the RIES the corresponding weight coefficient isgiven to different energy output types For an RIES theuncertainties of the source-side wind light and otheruncertain resources are large and have a great influence onthe stable operation of the system and the energy dispatchmanagement +erefore the source-side wind forecastingis set to a higher weight in addition as the power gridplays a leading role in the integrated energy system byvirtue of its perfect architecture and central hub advan-tages correspondingly the power load forecast is set to ahigher weight

+is paper proposes three calculation indicators namelymean absolute percentage error (MAPE) mean accuracy

x1 x2 xm

x1 x2 xn

x1 x2 xk

Hiddenlayer

Multitask learningOutputlayer

Inputlayer

Single-task learning

Wind powerhistorical data

Electrical loadhistorical data

Multisourcehistorical data

Figure 5 Comparison of single-task and multitasking learningbased on the LSTM network

6 Complexity

(MA) and weightedmean accuracy (WMA) as shown in thefollowing equation

MAPE 1n

1113944

n

i1

P(i) minus 1113954P(i)

P(i)

11138681113868111386811138681113868111386811138681113868

11138681113868111386811138681113868111386811138681113868times 100

MA 1 minus MAPE

WMA λ1MA1 + λ2MA2 + middot middot middot + λkMAk

(3)

In the equation P(i) and 1113954P(i) respectively representthe actual energy value and the predicted energy value at thefirst moment n is the sample data amount MAk representsthe prediction accuracy of the k prediction task and λk

represents the weight coefficient of the k prediction task Inthis paper the weight ratios of the energy predictions onboth sides of the source and the load are set as shown inTable 1

6 Simulation Analysis

61 Data Description +e experimental data in this paperare derived from the measured data provided by RIES in acertain area of Tianjin +e RIES consists of a power systema thermal system and a natural gas system +e source sideincludes energy supply equipment such as wind turbinesand photovoltaic power generation equipment Couplinglinks include energy conversion equipment such as cold andthermal power supply equipment electric boilers and gasboilers Energy demand mainly includes power load heatload and cooling load +e load side and the energy storageside contain devices such as electrical energy storage andthermal energy storage to improve the flexibility of thesystem In addition in order to prove the validity of thesource-load vertical information interaction the historicaldata of the lighting load reflecting the typical electricity

usage behavior of the user are also sampled All historicalenergy data sampling interval is 1 h data from July 2015 toJune 2016 are used as training set data and data from July2016 are used as test set data with 1 h as time step based onthe historical data of the first 7 days the energy supply andenergy demand on both sides of the source and the load inthe next day are predicted +e total 57168 samples weredivided into two parts (92 as the train set and 8 as the testset) +e forecasting models were trained by the train datasetand validated on the test dataset +e statistical informationof the above dataset is shown in Table 2

62 Analysis of Examples +is section analyzes the effect ofRIES multienergy information interactive energy prediction+e selected sunny day and rainy day of the test set wereanalyzed separately +e energy prediction results of windpower output photovoltaic output electric load heat loadand cold load on both sides of the source and the load in thetwo scenarios are shown in Figures 7 and 8 respectively

+e prediction results of Figures 7 and 8 show that theproposed source-load parallel LSTM learning model canobtain ideal results in wind energy photovoltaic electricload thermal load and cold load energy prediction in sunnyday scenario and rainy day scenario which confirms theeffectiveness of the model proposed in this paper +eprediction effect of each subtask in the parallel predictionresults shows a large difference Compared with the windand light prediction on the source side the predicted curvesof the charge heat and cold load on the charge side arecloser to the actual sequence curve which is due to theintermittent and random nature of source-side wind powerand photovoltaics

Comparing Figures 7(c)ndash7(e) it can be found that theelectric load is more volatile than the thermal load and thecold load which is because the thermal inertia and cold

LSTM 1

LSTMn

Forecast 1input data

Forecast ninput data

Parameters sharing

Forecast 1output data

Forecast noutput data

Inputssharing

Use

r beh

avio

r

Mul

tiene

rgy

sour

ce-lo

ad in

tegr

ated

fore

casti

ng re

sults

Figure 6 Interactive energy prediction model of RIES information based on parallel LSTM

Complexity 7

inertia of the thermal system are stronger [39] and thevolatility is also even smaller +ermal inertia means that thedynamic change of the thermal system is a slow process froma time scale Compared with the power system the heatenergy transmission exhibits a delay effect and the upperand lower fluctuations of the heat load exhibit an inertialeffect and the cold inertia is similar Because of the existenceof thermal inertia and cold inertia the thermal load in thethermal system is less random in the short-term predicted

time scale showing a characteristic of smooth variation soits prediction uncertainty is also small and the thermal loadand the cold load are relative to the electricity +e pre-diction error of the load is also relatively small ComparingFigures 7 and 8 it can be found that the uncertainty of actualphotovoltaic power and actual wind power in rainy dayscenes is larger than that of sunny days+e forecast curve ofphotovoltaic and wind power in sunny scenes is closer to thetrue value the electric load and heat load in rainy scenes

Table 1 Weight settings of source-load double-sided energy prediction tasks

Prediction task Wind power PV power Electric load Heat load Cold loadWeight coefficient 025 025 02 015 015

Table 2 +e description of simulation dataset

Data type Max (kW) Median (kW) Min (kW) Mean (kW) Std (kW) Train set size Test set sizeWind power 317561 233774 36612 216102 78964 8784 744Photovoltaic power 235000 0000 0000 42507 70557 8784 744Electric load 696585 584923 209872 523803 139174 8784 744Heat load 148473 92719 60532 100076 27240 8784 744Cold load 436964 223213 39046 221949 135304 8784 744Lighting load 03333 50000 2778 44792 27734 8784 744Multisource fusion mdash mdash mdash mdash mdash 52704 4464

True valuePredictive value

50

100

150

200

250

300

Win

d po

wer

(kW

)

5 10 15 20 250t (h)

(a)

True valuePredictive value

5 10 15 20 250t (h)

0

50

100

150

200

250

PV p

ower

(kW

)

(b)

True valuePredictive value

5 10 15 20 250t (h)

200

300

400

500

600

700

Elec

tric

load

(kW

)

(c)

True valuePredictive value

5 10 15 20 250t (h)

40

60

80

100

120

140

Hea

d lo

ad (k

W)

(d)

True valuePredictive value

5 10 15 20 250t (h)

0

100

200

300

400

500

Col

d lo

ad (k

W)

(e)

Figure 7 +e results of source-load bilateral energy prediction within a sunny day (a) Prediction of wind power (b) Prediction of PVpower (c) Prediction of electric load (d) Prediction of heat load (e) Prediction of cold load

8 Complexity

should increase as a whole +is is due to the cold weather inthe rainy scenes which affects the userrsquos energy use behaviorand increases the userrsquos demand for electricity and heat thusincreasing the electrical load and heat load

621 Comparative Analysis of Multienergy Load InformationInteraction In order to verify the effect of interactivelearning of horizontal and vertical multienergy informationthe parallel LSTM predictions considering user energy be-havior and parallel LSTM prediction without consideringuser energy behavior and the results of single LSTM pre-diction without considering user energy behavior are pro-posed to conduct a comparative analysis In order to ensurethe fairness of the experimental test this paper uses the samedataset for training and testing

+e comparison experiment setup is shown in Table 3 Inthe parallel LSTM learning without considering the userrsquosenergy behavior the input sample is based on the originalprediction model A and the lighting load data reflecting theuser behavior in the input sample are not considered +eprediction is based only on the historical data of the pre-dicted object the model C does not consider the singleLSTMprediction of the userrsquos energy behavior and the inputand output of each subprediction model only consider theinput and output of a single energy and there is also noparameter sharing between the LSTM unit of the different

submodels Model B uses parallel LSTM learning +ere isparameter sharing among each subtask which belongs to amultitask learning Model C uses a single LSTM learningand each subtask does not affect each other It belongs to asingle-task learning

+e prediction accuracy of the multienergy charge in-formation interaction comparison experiment is shown inTable 4 and Figure 9 +rough the analysis of Table 4 andFigure 9 it shows that the average prediction accuracy of theweights of the models A B and C are 09009 08917 and08728 respectively +e prediction result of the model A isthe best the model B is the second and the model C is thethird +at is considering the interaction of vertical andhorizontal information the energy predictions on both sidesof the source and the load show better prediction accuracy

Comparing the prediction results of Model B and ModelC for the five subprediction tasks when the different qualityenergies are predicted together in parallel a higher precisionthan the single prediction is obtained It can be seen thatparallel LSTM learning considering horizontal multienergyinformation interaction utilizes the coupled data betweendifferent systems to fully exploit the interactive informationbetween different quality energy sources of the integratedenergy system so that the prediction performance of acertain energy can be improved In addition considering themultienergy horizontal information interaction the pre-dicted accuracy of the load-side electrical load heat load and

True valuePredictive value

5 10 15 20 250t (h)

100

150

200

250

300

350W

ind

pow

er (k

W)

(a)

True valuePredictive value

5 10 15 20 250t (h)

0

20

40

60

80

100

120

PV p

ower

(kW

)(b)

True valuePredictive value

5 10 15 20 250t (h)

300

400

500

600

700

Elec

tric

load

(kW

)

(c)

True valuePredictive value

5 10 15 20 250t (h)

60

80

100

120

140

160

Hea

t loa

d (k

W)

(d)

True valuePredictive value

5 10 15 20 250t (h)

0

100

200

300

400

500

Col

d lo

ad (k

W)

(e)

Figure 8+e results of source-load bilateral energy prediction within a rainy day (a) Prediction of wind power (b) Prediction of PV power(c) Prediction of electric load (d) Prediction of heat load (e) Prediction of cold load

Complexity 9

cold load are increased by 313 253 and 267 re-spectively while the predicted accuracy of the source-sidewind power and photovoltaic is increased by 152 and146 and the energy prediction accuracy of the charge sideis relatively high It can be seen that the source-side un-certainty resources are also affected by other energy in-teractions in the system but the impact is small while thecharge side demand is largely affected by other energy in-teractions in the system And the information interaction ismore intimate between each other

Comparing the prediction results of Model A and ModelB when the userrsquos energy behavior information interactsbetween the source and the load in vertical direction theaverage weight accuracy of the overall prediction is increasedby 103 Comparing the performance improvement effectsof the five subprediction tasks it can be found that thepercentage improvement of PV prediction accuracy is185 and the improvement effect is relatively best fol-lowed by the electrical load while the prediction accuracy ofwind power heat load and cooling load is not obvious Itcan be seen that considering the source-load vertical in-formation interaction can improve the energy predictionaccuracy on both sides of the source and the load but thespecific user energy behavior has different differences in theenergy prediction of different types of users and the user

lighting behavior is obviously related to photovoltaic powergeneration and user electricity +e interaction betweenloads is more tight and its effect on PV prediction andelectric load forecasting is more obvious

622 Comparison of Deep Neural Network Algorithms withOther Algorithms In order to verify the prediction effect ofthe LSTM deep neural network algorithm the LSTM modeland the autoregressive integrated moving average model(ARIMA) algorithm and the classic BP neural network(BPNN) algorithm in this paper are proposed +e simu-lation results of them were compared and analyzed In orderto ensure the fairness of the experimental test this paper usesthe same dataset for training and testing and both use asingle-task prediction strategy +e prediction results of thethree algorithms are shown in Table 5 and Figure 10

Observing the prediction results the prediction effect ofARIMA algorithm is the most general and the predictionaccuracy of BPNN algorithm is much higher than that ofARIMA algorithm +e prediction accuracy of each task ofLSTM deep learning prediction algorithm proposed in thispaper is the highest +e ARIMA algorithm can not effec-tively use the historical information in the long-term se-quence because it only uses a part of the historicalinformation to train which leads to the large predictionerror +e LSTM algorithm can fully utilize its strongnonlinear learning ability and long-term and short-termmemory advantages to exploit the information value of long-term historical data so the grasp of the energy prediction lawis more precise

In addition comparing the energy prediction results ofthe LSTM algorithm and the BPNN algorithm one by onethe LSTM algorithm predicts that the accuracy of thethermal load and the cold load prediction is more obvious Itcan be seen that due to the greater inertia of the thermalsystem the time scale of the change law is longer whichmakes the prediction accuracy of the long-term and short-term deep learning on the heat load and the cold loadsignificantly higher than the electric load

In order to evaluate the computational efficiency ofalgorithms four forecasting scenarios are performed tentimes so that the forecasting results can reliably be sum-marized as in Table 6 Parallel LSTM and Single LSTM havelonger training time but the prediction time is kept within

Table 3 Contrast experiment setting of multisource source information interaction

Model Vertical interaction Horizontal interactionModel A User energy behavior is considered Parallel LSTM learningModel B User energy behavior is not considered Parallel LSTM learningModel C User energy behavior is not considered Single-LSTM learning

Table 4 +ree kinds of model prediction results of information interaction comparison experiment

Precision Wind power PV power Electric load Heat load Cold load Mean precisionModel A 08602 08735 09236 09445 09407 09009Model B 08542 08576 09113 09401 09368 08917Model C 08414 08453 08836 09169 09124 08728

Model AModel BModel C

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

PV Electric Heat Cold MeanWindPrediction task

Figure 9 Comparison of prediction effects of three models

10 Complexity

one minute and can meet the requirements of short-termforecasting Further analysis shows that the parallel trainingtime and prediction time of parallel LSTM is greatly im-proved compared to single LSTM due to the combination offive subtasks into one multitask learning In summary al-though the LSTM algorithm is time consuming in trainingtime for short-term prediction the training process doesnot require online completion and the prediction time ofthe LSTM algorithm especially parallel LSTM can fullymeet the requirements of short-term prediction Consid-ering the prediction accuracy the sacrifice of the LSTMalgorithm in computing time is worthwhile

7 Conclusion

In this paper the horizontal and vertical information interactioncharacteristics between multiple energy loads in RIES areconsidered Using LSTM network deep learning algorithm a

parallel LSTM-based multienergy-load information interactiveshort-term energy prediction method is proposed+rough thesimulation of the test data the following conclusions are drawn

(1) +e ldquoinformation interactionrdquo strategy proposed inthis paper can effectively improve the predictionperformance of source-load energy prediction inregional integrated energy systems

(2) Parallel prediction considering horizontal in-teraction mainly affects the prediction effect on theload side while the information sharing of user il-lumination behavior considering vertical interactionmainly affects the improvement of PV predictioneffect

(3) Compared with the conventional prediction algo-rithm the LSTM deep learning algorithm can learnthe deep hidden information of historical dataimprove the prediction accuracy while meetingcalculation time and prove the effectiveness of thealgorithm in the prediction of time series featuresequence

However the performance of the proposed parallelLSTM algorithm may be further enhanced by some evo-lutionary algorithms or multiobjective optimization algo-rithms such as differential evolution algorithm [40]multiobjective extremal optimization [41] and evolutionarymultiobjective optimization algorithms [42ndash47] Of coursesome significant topics are worth researching on the opti-mization of the proposed parallel LSTM by multiobjectiveoptimization

Data Availability

+e data used to support the findings of this study areavailable from the corresponding author upon request

Conflicts of Interest

+e authors declare no potential conflicts of interest

Acknowledgments

+is work was fully supported by the National Key Researchand Development Program of China (no 2018YFB0904200)

References

[1] A Q Huang M L Crow G T Heydt J P Zheng andS J Dale ldquo+e future renewable electric energy delivery andmanagement (FREEDM) system the energy internetrdquo Pro-ceedings of the IEEE vol 99 no 1 pp 133ndash148 2011

[2] I Colak S Sagiroglu G Fulli M Yesilbudak and C CovrigldquoA survey on the critical issues in smart grid technologiesrdquo

Table 5 Comparison of prediction results of three algorithms

Precision Wind power PV power Electric load Heat load Cold load Mean precisionLSTM 08414 08453 08836 09169 09124 08728BPNN 08216 08314 08695 08912 08889 08542ARMIA 07733 07824 08282 08454 08408 08075

Prediction precision ofwind powerPrediction precision ofPV powerPrediction precision ofelectric load

Prediction precision ofhead loadPrediction precision ofcold loadWeight average predictionprecision

07

075

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

BPNN ARMIALSTMPrediction algorithm

Figure 10 Comparison of prediction effects of three algorithms

Table 6 Comparison of computational efficiency of fouralgorithms

Algorithm Training timeaverage (sec)

Prediction timeaverage (sec)

Parallel LSTM 436374 1002Single LSTM 1218337 3662BPNN 1094911 2253ARMIA 54696 215

Complexity 11

Renewable and Sustainable Energy Reviews vol 54 pp 396ndash405 2016

[3] C Tu X He Z Shuai and F Jiang ldquoBig data issues in smartgrid-a reviewrdquo Renewable and Sustainable Energy Reviewsvol 79 pp 1099ndash1107 2017

[4] Y Zhang R Yang K Zhang H Jiang and J J ZhangldquoConsumption behavior analytics-aided energy forecastingand dispatchrdquo IEEE Intelligent Systems vol 32 no 4pp 59ndash63 2017

[5] J Deng F Wang Y Chen and X Zhao ldquoFrom industry 40 toenergy 50 the concept connotation and system framework ofintelligent energy systemsrdquo Journal of Automation vol 41no 12 pp 2003ndash2016 2015

[6] Q Sun L Shi Y Ni D Si and J Zhu ldquoAn enhanced cas-cading failure model integrating data mining techniquerdquoProtection and Control of Modern Power Systems vol 2 2017

[7] D Zheng A T Eseye J Zhang and H Li ldquoShort-term windpower forecasting using a double-stage hierarchical ANFISapproach for energy management in microgridsrdquo Protectionand Control of Modern Power Systems vol 2 pp 1ndash10 2017

[8] Z Li L Ye Y Zhao X Song J Teng and J Jin ldquoShort-termwind power prediction based on extreme learning machinewith error correctionrdquo Protection and Control of ModernPower Systems vol 1 no 1 pp 1ndash8 2016

[9] H Z Wang Z X Lei X Zhang B Zhou and J C Peng ldquoAreview of deep learning for renewable energy forecastingrdquoEnergy Conversion and Management vol 198 pp 1ndash16 2019

[10] H Z Wang G B Wang G Q Li J C Peng and Y T LiuldquoDeep belief network based deterministic and probabilisticwind speed forecasting approachrdquo Applied Energy vol 182pp 80ndash93 2016

[11] M-R Chen G-Q Zeng K-D Lu and J Weng ldquoA two-layernonlinear combination method for short-term wind speedprediction based on ELM ENN and LSTMrdquo IEEE Internet ofings Journal vol 6 no 4 pp 6997ndash7010 2019

[12] H Liu X Mi and Y Li ldquoSmart multi-step deep learningmodel for wind speed forecasting based on variational modedecomposition singular spectrum analysis LSTM networkand ELMrdquo Energy Conversion and Management vol 159pp 54ndash64 2018

[13] J Chen G-Q Zeng W Zhou W Du and K-D Lu ldquoWindspeed forecasting using nonlinear-learning ensemble of deeplearning time series prediction and extremal optimizationrdquoEnergy Conversion and Management vol 165 pp 681ndash6952018

[14] W Kong Z Y Dong Y Jia D J Hill Y Xu and Y ZhangldquoShort-term residential load forecasting based on LSTM re-current neural networkrdquo IEEE Transactions on Smart Gridvol 10 no 1 2017

[15] H Gao W Huang X Yang Y Duan and Y Yin ldquoTowardservice selection for workflow reconfiguration an interface-based computing solutionrdquo Future Generation ComputerSystems vol 87 pp 298ndash311 2018

[16] L Qi Q He F Chen et al ldquoFinding all you need web APIsrecommendation in web of things through keywords searchrdquoIEEE Transactions on Computational Social Systems vol 6no 5 2019

[17] L Qi R Wang C Hu S Li Q He and X Xu ldquoTime-awaredistributed service recommendation with privacy-preserva-tionrdquo Information Sciences vol 480 pp 354ndash364 2019

[18] Y Zhang and Q Yang ldquoAn overview of multi-task learningrdquoNational Science Review vol 5 no 1 pp 34ndash47 2018

[19] A Argyriou M T Evgeniou and M Pontil ldquoConvex multi-task feature learningrdquo Machine Learning vol 73 no 3pp 243ndash272 2008

[20] T Jebara ldquoMultitask sparsity via maximum entropy dis-criminationrdquo Journal of Machine Learning Research vol 12no 1 pp 75ndash110 2011

[21] L Qi Y Chen Y Yuan S Fu X Zhang and X Xu ldquoA QoS-aware virtual machine scheduling method for energy con-servation in cloud-based cyber-physical systemsrdquo WorldWide Web vol 5 pp 1ndash23 2019

[22] R Zhang P Xie C Wang G Liu and S Wan ldquoClassifyingtransportation mode and speed from trajectory data via deepmulti-scale learningrdquo Computer Networks vol 162 p 1068612019

[23] W Wang D Wang H Jia et al ldquoReview of steady-stateanalysis of typical regional integrated energy system under thebackground of energy internetrdquo Proceedings of the CSEE12pp 3292ndash3306 2016

[24] S Ding S Qu Y Xi and S Wan ldquoStimulus-driven andconcept-driven analysis for image caption generationrdquoNeurocomputing 2019

[25] P Vrba V Marik P Siano et al ldquoA review of agent andservice-oriented concepts applied to intelligent energy sys-temsrdquo IEEE Transactions on Industrial Informatics vol 10no 3 pp 1890ndash1903 2014

[26] W Jia C Kang and Q Chen ldquoAnalysis on demand-sideinteractive response capability for power system dispatch in asmart grid frameworkrdquo Electric Power Systems Researchvol 90 pp 11ndash17 2012

[27] Z C Lipton J Berkowitz and C Elkan ldquoA critical review ofrecurrent neural networks for sequence learningrdquo CoRR2015 httpsarxivorgpdf150600019pdf

[28] T N Sainath O Vinyals A Senior and H Sak ldquoCon-volutional long short-term memory fully connected deepneural networksrdquo in Proceedings of the 2015 IEEE In-ternational Conference on Acoustics Speech and Signal Pro-cessing (ICASSP) pp 4580ndash4584 IEEE Brisbane AustraliaApril 2015

[29] Y Guo Z Cheng L Nie YWang J Ma andM KankanhallildquoAttentive long short-term preference modeling for person-alized product searchrdquo ACM Transactions on InformationSystems (TOIS) vol 37 no 2 p 19 2019

[30] A Graves ldquoLong short-term memoryrdquo Neural Computationvol 9 pp 1735ndash1780 1997

[31] F A Gers J Schmidhuber and F Cummins ldquoLearning toforget continual prediction with LSTMrdquo Neural Computa-tion vol 12 no 10 pp 2451ndash2471 2000

[32] M L Seltzer and J Droppo ldquoMulti-task learning in deepneural networks for improved phoneme recognitionrdquo inProceedings of the 2013 IEEE International Conference onAcoustics Speech and Signal Processing pp 6965ndash6969 IEEEVancouver Canada May 2013

[33] S Wan Y Zhao T Wang Z Gu Q H Abbasi andK-K R Choo ldquoMulti-dimensional data indexing and rangequery processing via Voronoi diagram for internet of thingsrdquoFuture Generation Computer Systems vol 91 pp 382ndash3912019

[34] Z Gao D Y Wang S H Wan H Zhang and Y L WangldquoCognitive-inspired class-statistic matching with triple-con-strain for camera free 3D object retrievalrdquo Future GenerationComputer Systems vol 94 pp 641ndash653 2019

[35] C A Kang A R Brandt and L J Durlofsky ldquoOptimaloperation of an integrated energy system including fossil fuel

12 Complexity

power generation CO2 capture and windrdquo Energy vol 36no 12 pp 6806ndash6820 2011

[36] Q Zhang SWan BWang DW Gao and HMa ldquoAnomalydetection based on randommatrix theory for industrial powersystemsrdquo Journal of Systems Architecture vol 95 pp 67ndash742019

[37] Z Gao H Z Xuan H Zhang S Wan and K K R ChooldquoAdaptive fusion and category-level dictionary learningmodel for multi-view human action recognitionrdquo IEEE In-ternet of ings Journal p 1 2019

[38] WWu K Chen Y Qiao and Z Lu ldquoProbabilistic short-termwind power forecasting based on deep neural networksrdquo inProceedings of the 2016 International Conference on Proba-bilistic Methods Applied to Power Systems (PMAPS) pp 1ndash8Beijing China October 2016

[39] K M Powell A Sriprasad W J Cole and T F EdgarldquoHeating cooling and electrical load forecasting for a large-scale district energy systemrdquo Energy vol 74 pp 877ndash8852014

[40] A Peimankar S J Weddell T Jalal and A C LapthornldquoMulti-objective ensemble forecasting with an application topower transformersrdquo Applied Soft Computing vol 68pp 233ndash248 2018

[41] G-Q Zeng J Chen L-M Li et al ldquoAn improved multi-objective population-based extremal optimization algorithmwith polynomial mutationrdquo Information Sciences vol 330pp 49ndash73 2016

[42] Y-L Hu and L Chen ldquoA nonlinear hybrid wind speedforecasting model using LSTM network hysteretic ELM andDifferential Evolution algorithmrdquo Energy Conversion andManagement vol 173 pp 123ndash142 2018

[43] G-Q Zeng J Chen Y-X Dai L-M Li C-W Zheng andM-R Chen ldquoAn improved Design of fractional order PIDcontroller for automatic regulator voltage system based onmulti-objective extremal optimizationrdquo Neurocomputingvol 160 pp 173ndash184 2015

[44] X Xu R Gu F Dai L Qi and S Wan ldquoMulti-objectivecomputation offloading for internet of vehicles in cloud-edgecomputingrdquo Wireless Networks pp 1ndash19 2019

[45] S Wan X Li Y Xue W Lin and X Xu ldquoEfficient com-putation offloading for internet of vehicles in edge comput-ing-assisted 5G networksrdquo e Journal of Supercomputingpp 1ndash30 2019

[46] B Wang S Wan X Zhang and K-K R Choo ldquoA novelindex for assessing the robustness of integrated electricalnetwork and a natural gas networkrdquo IEEE Access vol 6pp 40400ndash40410 2018

[47] Y Yin F Yu Y Xu L Yu and J Mu ldquoNetwork location-aware service recommendation with random walk in cyber-physical systemsrdquo Sensors vol 17 no 9 p 2059 2017

Complexity 13

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Page 3: Parallel LSTM-Based Regional Integrated Energy System ...downloads.hindawi.com/journals/complexity/2019/7414318.pdf · Characteristics of Regional Integrated Energy System “Information

more tight [24] (1) From the perspective of horizontalinterconnection RIES will integrate the energy subsystemssuch as coldheatelectricgastraffic to construct a mul-tienergy flow system +e energy is converted and flowedbetween the systems through the coupling componentsand the various energy forms complement each other +eenergy resources have to be optimally configured in a widearea that is to achieve the ldquohorizontal information in-terconnectionrdquo of the energy Internet (2) From the per-spective of vertical interconnection based on thetraditional energy supply network RIES will realize thereal-time sharing of information resources in the source-network-load-storage through Internet technology +isresource sharing mechanism has two-way conductioncharacteristics +e user has information bidirectionalconduction capability between the information controlsystem the energy supply module and the multienergyintelligent delivery module +e interactive sharing ofenergy information can also be realized between users thatis to realize the ldquovertical information interconnectionrdquo ofthe energy Internet

3 RIES Multienergy CouplingCorrelation Analysis

31 RIES Horizontal Multienergy Information Coupling+e horizontal multienergy information coupling corre-lation of RIES is represented by the correlation betweendifferent energy flow subsystems such as electricity gasand heat

From the perspective of coupling the composition of theRIES can be divided into a coupling unit and a noncouplingunit As shown in Figure 2 the coupling unit connectsdifferent energy flow subsystems and consumes one or moreenergy sources to generate other energy sources Typicalcoupling forms are as follows cold and heat electricity

supply units consume coal or natural gas to generate elec-tricity heat and cold +e heat pump consumes electricalenergy to generate thermal energy the electrical hydrogenconsumes electrical energy to generate hydrogen and thecoupling unit causes the energy flow subsystems to interact+e uncoupled unit includes the network and equipmentinside each energy flow subsystem and other uncoupledboundary conditions such as source and load +e uncou-pled unit also affects other systems through the coupling unit[25]+erefore different energy flow subsystems have strongcoupling between them and there is a large amount ofhorizontal multienergy coupling information existing in themultienergy system data resulting in great correlation be-tween various units within the system

32 RIES Vertical Source-Load Information Coupling +evertical source-load information coupling correlation ofRIES is represented by the correlation between the multi-energy supply of the source and the userrsquos energy behavior of

WT

Input side

Distributionnetwork

Natural gas WindSolar

Output side Cold load Electric

loadHeat load

PV

Refrigeration

CCHPGB

ES

HS

Coupling side

Informationcommunicationsupport system

Calculationdata storage

layer

Informationcorrespondence

network

Acquisitionmonitoring

layer

Analysisdecision level

Lateral information interaction between multiplecapabilities

Vert

ical

info

rmat

ion

inte

ract

ion

betw

een

sour

ce lo

ads

sum

sum

Figure 1 Basic framework of RIES

Gas

Gasturbine

CHP

Powerto gas

Electric storage

Heatpump

Gasboiler

HeatCirculating pump

Heat storage

Gas storage

Electric

Figure 2 Transverse multicoupling schematic diagram

Complexity 3

the load Different energy behaviors correspond to differentload characteristics [26] For example the lighting behaviorin RIES is a typical user energy behavior +rough advancedmeasurement devices the userrsquos lighting behavior can bereflected as a lighting load curve

From the perspective of user energy behavior there aremultiple coupling relationships between multiple energyloads in RIES Taking the userrsquos lighting behavior as anexample if the cloud suddenly covers a certain area andreduces the ambient light intensity it will affect the userrsquoslighting behavior people will increase the lighting equip-ment usage to meet the lighting demand and the lightingload will increase At the same time such cloud movementswill also have an impact on photovoltaic power generationcloud clusters will block the photovoltaic panels which willaffect the irradiance reduction on the photovoltaic panelsand affect the photovoltaic power generation In order toprove the characteristics of this behavioral feature this papertakes the data of three days from January 24 to 26 2016 ofRIES in a certain area of Tianjin for quantitative analysis+ephotovoltaic power output curve and the user lighting loadcurve are shown in Figure 3

+e upper part of Figure 3 is the curve trend of thephotovoltaic power output and the lower part is the curvetrend of the userrsquos lighting load (1) Both the userrsquos lightingload and the photovoltaic power output have a day and nightcycle +e distribution of them is mainly concentrated in thedaytime +e peak of the userrsquos lighting load is generallydistributed between 18 00 and 20 00 and the peak outputof the photovoltaic power supply is generally distributedbetween 12 00 and 14 00 at noon (2) Taking the three-dayshort-term data for comparison analysis the overall outputof photovoltaic power generation on the third day wassignificantly larger than the previous two days Accordinglythe lighting load on the third day between 10 00 and 16 00is reduced compared to the previous two days and it isreduced to the low of the day It shows that when the outputof photovoltaic power generation increases it also affects theuserrsquos energy consumption behavior when the ambient lightintensity is large which means that the userrsquos demand forlighting energy is reduced (3) Take the ultra-short-termdata analysis from 12 00 to 16 00 on the first day andthe photovoltaic power output reaches the peak of one day at13 00 and the lighting load is also reduced to low of thedaytime at this moment During the period from 13 00 to14 00 a sudden downward climb occurs and the lightingload increases at the corresponding time +is ldquomutationrdquophenomenon should be caused by the cloudy weather af-fected by the cloud From the above analysis it can be foundthat there is a strong information coupling correlation be-tween the ldquosourcerdquo photovoltaic power output and theldquocharge endrdquo userrsquos lighting load

33 RIES Multienergy Information Interaction ModeRIES has both horizontal multienergy information couplingcorrelation characteristics and vertical source-load in-formation coupling correlation characteristics From theperspective of information interaction energy management

there are also two horizontal and vertical information in-teraction modes

+e horizontal information interaction mode is based onmultienergy conversion information sharing On the basis ofobtaining the measurement data of the multienergy flowsubsystem the nonlinear artificial intelligence deep learningtechnology is used to process the multienergy data in par-allel and it effectively utilizes the complex shared in-formation of energy conversion identifies the abstractfeatures of the training data and improves the precision ofsubtask prediction of the energy management

+e vertical information interaction mode is based onuser behavior information sharing On the basis of obtainingmulticategory data on both sides of the source and the loadthe deep learning method is used to dig the user-side be-havior characteristics and the associated user behaviorinformation data are taken as input to guide the source-sideenergy management prediction and improve the predictionaccuracy of the energy supply end

4 Multitask Learning Mechanism for DeepLearning Energy Prediction

41 LSTM Recurrent Neural Network

411 Principle of LSTM Cyclic Neural Network +e trainingobjects of the source-charged uncertain resources haveobvious timing characteristics and the sequence data havestrong correlation before and after From the perspective oflearning and training recurrent neural network (RNN) canprocess sequences with temporal correlation of arbitrarylength by using neurons with self-feedback which has ob-vious advantages in processing sequence data [27] +estructure of the RNN is shown in Figure 4(a) +e LSTMnetwork is a variant of RNN that overcomes the traditionalRNN gradient demise problem [28] +e hidden layer is nolonger a simple neural unit but an LSTM unit with a uniquememory mode Each LSTM unit contains a state cell thatdescribes the current state of the LSTM unit [29] c repre-sents the current state quantity of the LSTM unit and h

represents the current output of the LSTM unitLSTM reads and modifies state units by controlling the

forget gates input gates and output gates [30] which are

0

ndash100

250

50

100

150

200

ndash50

ndash200

ndash100

0

100

200

300

P (k

W)

10 20 30 40 50 60 700t (h)

Figure 3 +e relationship between the photovoltaic power curveand the lighting load curve of the user

4 Complexity

typically formed by sigmoid or tanh functions and Hada-mard product operation As shown in Figure 4(b) the forgetgate determines ctminus 1 of the previous moment and how manycomponents of the cell state ct remain at the current timeand the input gate determines how many components of thenetwork input are saved to the cell state ct at the current timext and howmany components of the output gate control cellstate ct are output to the current value ht of the LSTM

+e calculation formula between the variables of theLSTM unit is as follows

ft σ Wfxxi + Wfhhtminus 1 + bf1113872 1113873

it σ Wixxi + Wihhtminus 1 + bi( 1113857

gt ϕ Wgxxt + Wghhtminus 1 + bg1113872 1113873

ct ft ⊙ ctminus 1 + it ⊙gt

ot σ Woxxt + Wohhtminus 1 + bo( 1113857

ht ot ⊙ϕ ct( 1113857

(1)

In the formulaWfxWfhWixWihWgxWghWox andWoh are the weight matrices corresponding to the input of thenetwork activation function bf bi bg and bo are the offsetvectors ⊙ represents the multiplication of the Hadamardproduct matrix σ represents the sigmoid activation functionand ϕ represents the tanh activation function

412 LSTMDeep Learning Training Algorithm +e trainingof LSTM network adopts BP-based backpropagationthrough time (BPTT) [31] +e algorithm steps are as fol-lows (1) Calculate the output value of each neuron forwardFor the LSTMunit the values of the six vectors such asft itgt ct ot and ht are calculated (2) Calculate the error termof each neuron in reverse Similar to traditional RNN thebackpropagation of the LSTM error term consists of twolevels one is at the spatial level and the error term ispropagated to the upper layer of the network the other is atthe time level which propagates back in time For examplefrom the current time of t the error at each moment iscalculated (3) Calculate the gradient of each weightaccording to the corresponding error term and update thenetwork weight parameter (4) Jump to the first step andcontinue to perform the first 3 steps until the network erroris less than the given value

42 LSTM Network Multitask Learning MechanismMultitask learning (MTL) as an inductive migrationmechanism avoids the underfitting of model training byusing the correlation between multiple tasks digging thehidden common features and improving the generalizationlearning ability of the model [32] RIES source wind andlight prediction and load-end electricity gas heat and coldload prediction can be regarded as different tasks in mul-titask learning [33] Multiple tasks share some common dataor model structures that is there is a certain correlationbetween the multienergy prediction tasks +e multitasklearning prediction model based on the LSTM networkproposed in this paper shares the hidden layer nodes amongall prediction tasks through the hard parameter sharingmethod and trains the training data corresponding to themultienergy prediction tasks at the same time consideringthe correlation information between tasks to improve thelearning ability of all predictive tasks [34] A comparison ofsingle-task and multitask learning based on the LSTMnetwork is shown in Figure 5 +e multisource historicaldata in the figure is a collection of five types of historical dataincluding wind power photovoltaic electric load heat loadand cold load

5 Parallel LSTM-Based RIES InformationInteractive Energy Prediction Method

51 RIES Multienergy Prediction Model Design RIES hasuncertainties at both the source and the load end so accurateprediction of uncertain resources is the basis for imple-menting RIES energy optimization management andscheduling [35] +e source mainly includes wind powerprediction and photovoltaic power prediction and the loadmainly includes electric load prediction thermal load pre-diction and cold load prediction

As described in the second section of this paper there areintricate mutual coupling relationships between multipleenergy loads of RIES and the prediction tasks of uncertainresources on both sides of the source and load have strongcorrelation It can effectively improve the overall predictionaccuracy of the RIES uncertain resources by learning thiscoupling relationship through joint prediction training [36]From the perspective of RIES multienergy-load informationinteraction this paper proposes the concept of ldquoparallel

xtndash1 xt xt+1

htndash2

ctndash2

hh h

cc c

htndash1 ht ht+1

ctndash1 ct ct+1

(a)

tanh

tanh

xt

ht

ctctndash1

Forgetgate

ft it ot

ct

htndash1

Inputgate

Outputgate

σ σ σ

(b)

Figure 4 Schematic diagram of RNN and LSTM (a) +e structure of the RNN (b) +e structure of the LSTM cell

Complexity 5

predictionrdquo and utilizes the strong nonlinear learning abilityand historical memory advantages of LSTM deep learningnetwork to conduct integrated prediction of uncertain re-sources at both source and load As shown in Figure 6 thehorizontal information interaction based on multienergyconversion information sharingmainly utilizes themultitasklearning parameter sharing mechanism of the parallel LSTMprediction network the vertical information interactionbased on the user behavior information sharing mainly usesthe user behavior data to guide source-load multitask pre-diction as an additional unified input information Para-metric sharing is weight sharing which takes advantage ofthe correlation between subtasks Input sharing refers todata sharing Different subtasks share historical power datasuch as wind power photovoltaic power power loadthermal load and cooling load

52 Data Preprocessing +e training samples select the windpower generation time series on the source side the photo-voltaic power generation time series the charge side on the loadside the heat load the cold load time series and the illumi-nation load series data Considering the input and output rangeof the nonlinear activation function in the model in order toavoid neuron saturation the data are normalized and mappedbetween [minus 1 1] as shown in the following equation

xprime x minus xmax + xmin( 11138572

xmax minus xmin( 11138572 (2)

In the equation x is the actual input or output data andxmax and xmin are the maximum and minimum values of thevariable respectively

53 Task Construction +e essence of using the multitasklearning method for multienergy-load integration predictionlies in the close correlation of each prediction subtask [37]+e traditional single-task prediction method uses a singledimension of historical observations to predict its futureoutput and the model output is a single attribute +emultitask learning integration prediction uses the multidi-mensional data on both sides of the source and the load as a

common input to effectively utilize the sample On the otherhand through the shared hidden layer node divisionmethodthe five attributes of wind photovoltaic power generationelectricity heat and cold load are simultaneously predictedto play the role of inductive bias and improve predictionaccuracy

54 Model Training +e LSTM-based source-charge in-tegrated prediction model mainly needs to determine thefollowing parameters input layer dimension input layertime step hidden layer number each hidden layer di-mension and output layer dimension [38]

+e time step of the input layer is generally the length ofthe variable time series used for the source-load integrationprediction According to the periodic characteristics of theprediction object sequence the input layer time step is set to24 that is the historical data whose resolution ratio is 1 hduring the first seven days before the input is used forforecasting and it takes up one input node per day+e inputlayer dimension is the number of nodes +e input layernode of this paper contains six objects wind power gen-eration time series photovoltaic power generation timeseries charge side load heat load cold load time series andlighting load series which constitutes totally 42 nodes so theinput layer dimension is set to 42 Since this paper simul-taneously predicts six attributes of wind power photovol-taic electricity heat and cold load the output layer has 5nodes that is the output layer dimension is set to 5 In thispaper based on experience and multiple trial and com-parison the number of hidden layers is set to 2 and thedimension of the hidden layer is set to 30 +e selection ofthe activation function is as shown in Figure 4(a) thebatch_size is set to 42 the numbers of iterations of an epochis set to 8784 the learning rate is set to 080 and the learningdelay rate is set to 005 +e network training function usesthe BPTT algorithm mentioned above

55 Evaluation Indicators Multienergy-load integrationforecasting simultaneously trains multiple forecastingtasks To comprehensively measure the forecasting effect itis necessary to evaluate the integrated forecasting accuracyas a whole [39] In this paper the average accuracyevaluation index based on weight coefficient is proposedFor the importance degree of different energy on the twosides of the RIES the corresponding weight coefficient isgiven to different energy output types For an RIES theuncertainties of the source-side wind light and otheruncertain resources are large and have a great influence onthe stable operation of the system and the energy dispatchmanagement +erefore the source-side wind forecastingis set to a higher weight in addition as the power gridplays a leading role in the integrated energy system byvirtue of its perfect architecture and central hub advan-tages correspondingly the power load forecast is set to ahigher weight

+is paper proposes three calculation indicators namelymean absolute percentage error (MAPE) mean accuracy

x1 x2 xm

x1 x2 xn

x1 x2 xk

Hiddenlayer

Multitask learningOutputlayer

Inputlayer

Single-task learning

Wind powerhistorical data

Electrical loadhistorical data

Multisourcehistorical data

Figure 5 Comparison of single-task and multitasking learningbased on the LSTM network

6 Complexity

(MA) and weightedmean accuracy (WMA) as shown in thefollowing equation

MAPE 1n

1113944

n

i1

P(i) minus 1113954P(i)

P(i)

11138681113868111386811138681113868111386811138681113868

11138681113868111386811138681113868111386811138681113868times 100

MA 1 minus MAPE

WMA λ1MA1 + λ2MA2 + middot middot middot + λkMAk

(3)

In the equation P(i) and 1113954P(i) respectively representthe actual energy value and the predicted energy value at thefirst moment n is the sample data amount MAk representsthe prediction accuracy of the k prediction task and λk

represents the weight coefficient of the k prediction task Inthis paper the weight ratios of the energy predictions onboth sides of the source and the load are set as shown inTable 1

6 Simulation Analysis

61 Data Description +e experimental data in this paperare derived from the measured data provided by RIES in acertain area of Tianjin +e RIES consists of a power systema thermal system and a natural gas system +e source sideincludes energy supply equipment such as wind turbinesand photovoltaic power generation equipment Couplinglinks include energy conversion equipment such as cold andthermal power supply equipment electric boilers and gasboilers Energy demand mainly includes power load heatload and cooling load +e load side and the energy storageside contain devices such as electrical energy storage andthermal energy storage to improve the flexibility of thesystem In addition in order to prove the validity of thesource-load vertical information interaction the historicaldata of the lighting load reflecting the typical electricity

usage behavior of the user are also sampled All historicalenergy data sampling interval is 1 h data from July 2015 toJune 2016 are used as training set data and data from July2016 are used as test set data with 1 h as time step based onthe historical data of the first 7 days the energy supply andenergy demand on both sides of the source and the load inthe next day are predicted +e total 57168 samples weredivided into two parts (92 as the train set and 8 as the testset) +e forecasting models were trained by the train datasetand validated on the test dataset +e statistical informationof the above dataset is shown in Table 2

62 Analysis of Examples +is section analyzes the effect ofRIES multienergy information interactive energy prediction+e selected sunny day and rainy day of the test set wereanalyzed separately +e energy prediction results of windpower output photovoltaic output electric load heat loadand cold load on both sides of the source and the load in thetwo scenarios are shown in Figures 7 and 8 respectively

+e prediction results of Figures 7 and 8 show that theproposed source-load parallel LSTM learning model canobtain ideal results in wind energy photovoltaic electricload thermal load and cold load energy prediction in sunnyday scenario and rainy day scenario which confirms theeffectiveness of the model proposed in this paper +eprediction effect of each subtask in the parallel predictionresults shows a large difference Compared with the windand light prediction on the source side the predicted curvesof the charge heat and cold load on the charge side arecloser to the actual sequence curve which is due to theintermittent and random nature of source-side wind powerand photovoltaics

Comparing Figures 7(c)ndash7(e) it can be found that theelectric load is more volatile than the thermal load and thecold load which is because the thermal inertia and cold

LSTM 1

LSTMn

Forecast 1input data

Forecast ninput data

Parameters sharing

Forecast 1output data

Forecast noutput data

Inputssharing

Use

r beh

avio

r

Mul

tiene

rgy

sour

ce-lo

ad in

tegr

ated

fore

casti

ng re

sults

Figure 6 Interactive energy prediction model of RIES information based on parallel LSTM

Complexity 7

inertia of the thermal system are stronger [39] and thevolatility is also even smaller +ermal inertia means that thedynamic change of the thermal system is a slow process froma time scale Compared with the power system the heatenergy transmission exhibits a delay effect and the upperand lower fluctuations of the heat load exhibit an inertialeffect and the cold inertia is similar Because of the existenceof thermal inertia and cold inertia the thermal load in thethermal system is less random in the short-term predicted

time scale showing a characteristic of smooth variation soits prediction uncertainty is also small and the thermal loadand the cold load are relative to the electricity +e pre-diction error of the load is also relatively small ComparingFigures 7 and 8 it can be found that the uncertainty of actualphotovoltaic power and actual wind power in rainy dayscenes is larger than that of sunny days+e forecast curve ofphotovoltaic and wind power in sunny scenes is closer to thetrue value the electric load and heat load in rainy scenes

Table 1 Weight settings of source-load double-sided energy prediction tasks

Prediction task Wind power PV power Electric load Heat load Cold loadWeight coefficient 025 025 02 015 015

Table 2 +e description of simulation dataset

Data type Max (kW) Median (kW) Min (kW) Mean (kW) Std (kW) Train set size Test set sizeWind power 317561 233774 36612 216102 78964 8784 744Photovoltaic power 235000 0000 0000 42507 70557 8784 744Electric load 696585 584923 209872 523803 139174 8784 744Heat load 148473 92719 60532 100076 27240 8784 744Cold load 436964 223213 39046 221949 135304 8784 744Lighting load 03333 50000 2778 44792 27734 8784 744Multisource fusion mdash mdash mdash mdash mdash 52704 4464

True valuePredictive value

50

100

150

200

250

300

Win

d po

wer

(kW

)

5 10 15 20 250t (h)

(a)

True valuePredictive value

5 10 15 20 250t (h)

0

50

100

150

200

250

PV p

ower

(kW

)

(b)

True valuePredictive value

5 10 15 20 250t (h)

200

300

400

500

600

700

Elec

tric

load

(kW

)

(c)

True valuePredictive value

5 10 15 20 250t (h)

40

60

80

100

120

140

Hea

d lo

ad (k

W)

(d)

True valuePredictive value

5 10 15 20 250t (h)

0

100

200

300

400

500

Col

d lo

ad (k

W)

(e)

Figure 7 +e results of source-load bilateral energy prediction within a sunny day (a) Prediction of wind power (b) Prediction of PVpower (c) Prediction of electric load (d) Prediction of heat load (e) Prediction of cold load

8 Complexity

should increase as a whole +is is due to the cold weather inthe rainy scenes which affects the userrsquos energy use behaviorand increases the userrsquos demand for electricity and heat thusincreasing the electrical load and heat load

621 Comparative Analysis of Multienergy Load InformationInteraction In order to verify the effect of interactivelearning of horizontal and vertical multienergy informationthe parallel LSTM predictions considering user energy be-havior and parallel LSTM prediction without consideringuser energy behavior and the results of single LSTM pre-diction without considering user energy behavior are pro-posed to conduct a comparative analysis In order to ensurethe fairness of the experimental test this paper uses the samedataset for training and testing

+e comparison experiment setup is shown in Table 3 Inthe parallel LSTM learning without considering the userrsquosenergy behavior the input sample is based on the originalprediction model A and the lighting load data reflecting theuser behavior in the input sample are not considered +eprediction is based only on the historical data of the pre-dicted object the model C does not consider the singleLSTMprediction of the userrsquos energy behavior and the inputand output of each subprediction model only consider theinput and output of a single energy and there is also noparameter sharing between the LSTM unit of the different

submodels Model B uses parallel LSTM learning +ere isparameter sharing among each subtask which belongs to amultitask learning Model C uses a single LSTM learningand each subtask does not affect each other It belongs to asingle-task learning

+e prediction accuracy of the multienergy charge in-formation interaction comparison experiment is shown inTable 4 and Figure 9 +rough the analysis of Table 4 andFigure 9 it shows that the average prediction accuracy of theweights of the models A B and C are 09009 08917 and08728 respectively +e prediction result of the model A isthe best the model B is the second and the model C is thethird +at is considering the interaction of vertical andhorizontal information the energy predictions on both sidesof the source and the load show better prediction accuracy

Comparing the prediction results of Model B and ModelC for the five subprediction tasks when the different qualityenergies are predicted together in parallel a higher precisionthan the single prediction is obtained It can be seen thatparallel LSTM learning considering horizontal multienergyinformation interaction utilizes the coupled data betweendifferent systems to fully exploit the interactive informationbetween different quality energy sources of the integratedenergy system so that the prediction performance of acertain energy can be improved In addition considering themultienergy horizontal information interaction the pre-dicted accuracy of the load-side electrical load heat load and

True valuePredictive value

5 10 15 20 250t (h)

100

150

200

250

300

350W

ind

pow

er (k

W)

(a)

True valuePredictive value

5 10 15 20 250t (h)

0

20

40

60

80

100

120

PV p

ower

(kW

)(b)

True valuePredictive value

5 10 15 20 250t (h)

300

400

500

600

700

Elec

tric

load

(kW

)

(c)

True valuePredictive value

5 10 15 20 250t (h)

60

80

100

120

140

160

Hea

t loa

d (k

W)

(d)

True valuePredictive value

5 10 15 20 250t (h)

0

100

200

300

400

500

Col

d lo

ad (k

W)

(e)

Figure 8+e results of source-load bilateral energy prediction within a rainy day (a) Prediction of wind power (b) Prediction of PV power(c) Prediction of electric load (d) Prediction of heat load (e) Prediction of cold load

Complexity 9

cold load are increased by 313 253 and 267 re-spectively while the predicted accuracy of the source-sidewind power and photovoltaic is increased by 152 and146 and the energy prediction accuracy of the charge sideis relatively high It can be seen that the source-side un-certainty resources are also affected by other energy in-teractions in the system but the impact is small while thecharge side demand is largely affected by other energy in-teractions in the system And the information interaction ismore intimate between each other

Comparing the prediction results of Model A and ModelB when the userrsquos energy behavior information interactsbetween the source and the load in vertical direction theaverage weight accuracy of the overall prediction is increasedby 103 Comparing the performance improvement effectsof the five subprediction tasks it can be found that thepercentage improvement of PV prediction accuracy is185 and the improvement effect is relatively best fol-lowed by the electrical load while the prediction accuracy ofwind power heat load and cooling load is not obvious Itcan be seen that considering the source-load vertical in-formation interaction can improve the energy predictionaccuracy on both sides of the source and the load but thespecific user energy behavior has different differences in theenergy prediction of different types of users and the user

lighting behavior is obviously related to photovoltaic powergeneration and user electricity +e interaction betweenloads is more tight and its effect on PV prediction andelectric load forecasting is more obvious

622 Comparison of Deep Neural Network Algorithms withOther Algorithms In order to verify the prediction effect ofthe LSTM deep neural network algorithm the LSTM modeland the autoregressive integrated moving average model(ARIMA) algorithm and the classic BP neural network(BPNN) algorithm in this paper are proposed +e simu-lation results of them were compared and analyzed In orderto ensure the fairness of the experimental test this paper usesthe same dataset for training and testing and both use asingle-task prediction strategy +e prediction results of thethree algorithms are shown in Table 5 and Figure 10

Observing the prediction results the prediction effect ofARIMA algorithm is the most general and the predictionaccuracy of BPNN algorithm is much higher than that ofARIMA algorithm +e prediction accuracy of each task ofLSTM deep learning prediction algorithm proposed in thispaper is the highest +e ARIMA algorithm can not effec-tively use the historical information in the long-term se-quence because it only uses a part of the historicalinformation to train which leads to the large predictionerror +e LSTM algorithm can fully utilize its strongnonlinear learning ability and long-term and short-termmemory advantages to exploit the information value of long-term historical data so the grasp of the energy prediction lawis more precise

In addition comparing the energy prediction results ofthe LSTM algorithm and the BPNN algorithm one by onethe LSTM algorithm predicts that the accuracy of thethermal load and the cold load prediction is more obvious Itcan be seen that due to the greater inertia of the thermalsystem the time scale of the change law is longer whichmakes the prediction accuracy of the long-term and short-term deep learning on the heat load and the cold loadsignificantly higher than the electric load

In order to evaluate the computational efficiency ofalgorithms four forecasting scenarios are performed tentimes so that the forecasting results can reliably be sum-marized as in Table 6 Parallel LSTM and Single LSTM havelonger training time but the prediction time is kept within

Table 3 Contrast experiment setting of multisource source information interaction

Model Vertical interaction Horizontal interactionModel A User energy behavior is considered Parallel LSTM learningModel B User energy behavior is not considered Parallel LSTM learningModel C User energy behavior is not considered Single-LSTM learning

Table 4 +ree kinds of model prediction results of information interaction comparison experiment

Precision Wind power PV power Electric load Heat load Cold load Mean precisionModel A 08602 08735 09236 09445 09407 09009Model B 08542 08576 09113 09401 09368 08917Model C 08414 08453 08836 09169 09124 08728

Model AModel BModel C

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

PV Electric Heat Cold MeanWindPrediction task

Figure 9 Comparison of prediction effects of three models

10 Complexity

one minute and can meet the requirements of short-termforecasting Further analysis shows that the parallel trainingtime and prediction time of parallel LSTM is greatly im-proved compared to single LSTM due to the combination offive subtasks into one multitask learning In summary al-though the LSTM algorithm is time consuming in trainingtime for short-term prediction the training process doesnot require online completion and the prediction time ofthe LSTM algorithm especially parallel LSTM can fullymeet the requirements of short-term prediction Consid-ering the prediction accuracy the sacrifice of the LSTMalgorithm in computing time is worthwhile

7 Conclusion

In this paper the horizontal and vertical information interactioncharacteristics between multiple energy loads in RIES areconsidered Using LSTM network deep learning algorithm a

parallel LSTM-based multienergy-load information interactiveshort-term energy prediction method is proposed+rough thesimulation of the test data the following conclusions are drawn

(1) +e ldquoinformation interactionrdquo strategy proposed inthis paper can effectively improve the predictionperformance of source-load energy prediction inregional integrated energy systems

(2) Parallel prediction considering horizontal in-teraction mainly affects the prediction effect on theload side while the information sharing of user il-lumination behavior considering vertical interactionmainly affects the improvement of PV predictioneffect

(3) Compared with the conventional prediction algo-rithm the LSTM deep learning algorithm can learnthe deep hidden information of historical dataimprove the prediction accuracy while meetingcalculation time and prove the effectiveness of thealgorithm in the prediction of time series featuresequence

However the performance of the proposed parallelLSTM algorithm may be further enhanced by some evo-lutionary algorithms or multiobjective optimization algo-rithms such as differential evolution algorithm [40]multiobjective extremal optimization [41] and evolutionarymultiobjective optimization algorithms [42ndash47] Of coursesome significant topics are worth researching on the opti-mization of the proposed parallel LSTM by multiobjectiveoptimization

Data Availability

+e data used to support the findings of this study areavailable from the corresponding author upon request

Conflicts of Interest

+e authors declare no potential conflicts of interest

Acknowledgments

+is work was fully supported by the National Key Researchand Development Program of China (no 2018YFB0904200)

References

[1] A Q Huang M L Crow G T Heydt J P Zheng andS J Dale ldquo+e future renewable electric energy delivery andmanagement (FREEDM) system the energy internetrdquo Pro-ceedings of the IEEE vol 99 no 1 pp 133ndash148 2011

[2] I Colak S Sagiroglu G Fulli M Yesilbudak and C CovrigldquoA survey on the critical issues in smart grid technologiesrdquo

Table 5 Comparison of prediction results of three algorithms

Precision Wind power PV power Electric load Heat load Cold load Mean precisionLSTM 08414 08453 08836 09169 09124 08728BPNN 08216 08314 08695 08912 08889 08542ARMIA 07733 07824 08282 08454 08408 08075

Prediction precision ofwind powerPrediction precision ofPV powerPrediction precision ofelectric load

Prediction precision ofhead loadPrediction precision ofcold loadWeight average predictionprecision

07

075

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

BPNN ARMIALSTMPrediction algorithm

Figure 10 Comparison of prediction effects of three algorithms

Table 6 Comparison of computational efficiency of fouralgorithms

Algorithm Training timeaverage (sec)

Prediction timeaverage (sec)

Parallel LSTM 436374 1002Single LSTM 1218337 3662BPNN 1094911 2253ARMIA 54696 215

Complexity 11

Renewable and Sustainable Energy Reviews vol 54 pp 396ndash405 2016

[3] C Tu X He Z Shuai and F Jiang ldquoBig data issues in smartgrid-a reviewrdquo Renewable and Sustainable Energy Reviewsvol 79 pp 1099ndash1107 2017

[4] Y Zhang R Yang K Zhang H Jiang and J J ZhangldquoConsumption behavior analytics-aided energy forecastingand dispatchrdquo IEEE Intelligent Systems vol 32 no 4pp 59ndash63 2017

[5] J Deng F Wang Y Chen and X Zhao ldquoFrom industry 40 toenergy 50 the concept connotation and system framework ofintelligent energy systemsrdquo Journal of Automation vol 41no 12 pp 2003ndash2016 2015

[6] Q Sun L Shi Y Ni D Si and J Zhu ldquoAn enhanced cas-cading failure model integrating data mining techniquerdquoProtection and Control of Modern Power Systems vol 2 2017

[7] D Zheng A T Eseye J Zhang and H Li ldquoShort-term windpower forecasting using a double-stage hierarchical ANFISapproach for energy management in microgridsrdquo Protectionand Control of Modern Power Systems vol 2 pp 1ndash10 2017

[8] Z Li L Ye Y Zhao X Song J Teng and J Jin ldquoShort-termwind power prediction based on extreme learning machinewith error correctionrdquo Protection and Control of ModernPower Systems vol 1 no 1 pp 1ndash8 2016

[9] H Z Wang Z X Lei X Zhang B Zhou and J C Peng ldquoAreview of deep learning for renewable energy forecastingrdquoEnergy Conversion and Management vol 198 pp 1ndash16 2019

[10] H Z Wang G B Wang G Q Li J C Peng and Y T LiuldquoDeep belief network based deterministic and probabilisticwind speed forecasting approachrdquo Applied Energy vol 182pp 80ndash93 2016

[11] M-R Chen G-Q Zeng K-D Lu and J Weng ldquoA two-layernonlinear combination method for short-term wind speedprediction based on ELM ENN and LSTMrdquo IEEE Internet ofings Journal vol 6 no 4 pp 6997ndash7010 2019

[12] H Liu X Mi and Y Li ldquoSmart multi-step deep learningmodel for wind speed forecasting based on variational modedecomposition singular spectrum analysis LSTM networkand ELMrdquo Energy Conversion and Management vol 159pp 54ndash64 2018

[13] J Chen G-Q Zeng W Zhou W Du and K-D Lu ldquoWindspeed forecasting using nonlinear-learning ensemble of deeplearning time series prediction and extremal optimizationrdquoEnergy Conversion and Management vol 165 pp 681ndash6952018

[14] W Kong Z Y Dong Y Jia D J Hill Y Xu and Y ZhangldquoShort-term residential load forecasting based on LSTM re-current neural networkrdquo IEEE Transactions on Smart Gridvol 10 no 1 2017

[15] H Gao W Huang X Yang Y Duan and Y Yin ldquoTowardservice selection for workflow reconfiguration an interface-based computing solutionrdquo Future Generation ComputerSystems vol 87 pp 298ndash311 2018

[16] L Qi Q He F Chen et al ldquoFinding all you need web APIsrecommendation in web of things through keywords searchrdquoIEEE Transactions on Computational Social Systems vol 6no 5 2019

[17] L Qi R Wang C Hu S Li Q He and X Xu ldquoTime-awaredistributed service recommendation with privacy-preserva-tionrdquo Information Sciences vol 480 pp 354ndash364 2019

[18] Y Zhang and Q Yang ldquoAn overview of multi-task learningrdquoNational Science Review vol 5 no 1 pp 34ndash47 2018

[19] A Argyriou M T Evgeniou and M Pontil ldquoConvex multi-task feature learningrdquo Machine Learning vol 73 no 3pp 243ndash272 2008

[20] T Jebara ldquoMultitask sparsity via maximum entropy dis-criminationrdquo Journal of Machine Learning Research vol 12no 1 pp 75ndash110 2011

[21] L Qi Y Chen Y Yuan S Fu X Zhang and X Xu ldquoA QoS-aware virtual machine scheduling method for energy con-servation in cloud-based cyber-physical systemsrdquo WorldWide Web vol 5 pp 1ndash23 2019

[22] R Zhang P Xie C Wang G Liu and S Wan ldquoClassifyingtransportation mode and speed from trajectory data via deepmulti-scale learningrdquo Computer Networks vol 162 p 1068612019

[23] W Wang D Wang H Jia et al ldquoReview of steady-stateanalysis of typical regional integrated energy system under thebackground of energy internetrdquo Proceedings of the CSEE12pp 3292ndash3306 2016

[24] S Ding S Qu Y Xi and S Wan ldquoStimulus-driven andconcept-driven analysis for image caption generationrdquoNeurocomputing 2019

[25] P Vrba V Marik P Siano et al ldquoA review of agent andservice-oriented concepts applied to intelligent energy sys-temsrdquo IEEE Transactions on Industrial Informatics vol 10no 3 pp 1890ndash1903 2014

[26] W Jia C Kang and Q Chen ldquoAnalysis on demand-sideinteractive response capability for power system dispatch in asmart grid frameworkrdquo Electric Power Systems Researchvol 90 pp 11ndash17 2012

[27] Z C Lipton J Berkowitz and C Elkan ldquoA critical review ofrecurrent neural networks for sequence learningrdquo CoRR2015 httpsarxivorgpdf150600019pdf

[28] T N Sainath O Vinyals A Senior and H Sak ldquoCon-volutional long short-term memory fully connected deepneural networksrdquo in Proceedings of the 2015 IEEE In-ternational Conference on Acoustics Speech and Signal Pro-cessing (ICASSP) pp 4580ndash4584 IEEE Brisbane AustraliaApril 2015

[29] Y Guo Z Cheng L Nie YWang J Ma andM KankanhallildquoAttentive long short-term preference modeling for person-alized product searchrdquo ACM Transactions on InformationSystems (TOIS) vol 37 no 2 p 19 2019

[30] A Graves ldquoLong short-term memoryrdquo Neural Computationvol 9 pp 1735ndash1780 1997

[31] F A Gers J Schmidhuber and F Cummins ldquoLearning toforget continual prediction with LSTMrdquo Neural Computa-tion vol 12 no 10 pp 2451ndash2471 2000

[32] M L Seltzer and J Droppo ldquoMulti-task learning in deepneural networks for improved phoneme recognitionrdquo inProceedings of the 2013 IEEE International Conference onAcoustics Speech and Signal Processing pp 6965ndash6969 IEEEVancouver Canada May 2013

[33] S Wan Y Zhao T Wang Z Gu Q H Abbasi andK-K R Choo ldquoMulti-dimensional data indexing and rangequery processing via Voronoi diagram for internet of thingsrdquoFuture Generation Computer Systems vol 91 pp 382ndash3912019

[34] Z Gao D Y Wang S H Wan H Zhang and Y L WangldquoCognitive-inspired class-statistic matching with triple-con-strain for camera free 3D object retrievalrdquo Future GenerationComputer Systems vol 94 pp 641ndash653 2019

[35] C A Kang A R Brandt and L J Durlofsky ldquoOptimaloperation of an integrated energy system including fossil fuel

12 Complexity

power generation CO2 capture and windrdquo Energy vol 36no 12 pp 6806ndash6820 2011

[36] Q Zhang SWan BWang DW Gao and HMa ldquoAnomalydetection based on randommatrix theory for industrial powersystemsrdquo Journal of Systems Architecture vol 95 pp 67ndash742019

[37] Z Gao H Z Xuan H Zhang S Wan and K K R ChooldquoAdaptive fusion and category-level dictionary learningmodel for multi-view human action recognitionrdquo IEEE In-ternet of ings Journal p 1 2019

[38] WWu K Chen Y Qiao and Z Lu ldquoProbabilistic short-termwind power forecasting based on deep neural networksrdquo inProceedings of the 2016 International Conference on Proba-bilistic Methods Applied to Power Systems (PMAPS) pp 1ndash8Beijing China October 2016

[39] K M Powell A Sriprasad W J Cole and T F EdgarldquoHeating cooling and electrical load forecasting for a large-scale district energy systemrdquo Energy vol 74 pp 877ndash8852014

[40] A Peimankar S J Weddell T Jalal and A C LapthornldquoMulti-objective ensemble forecasting with an application topower transformersrdquo Applied Soft Computing vol 68pp 233ndash248 2018

[41] G-Q Zeng J Chen L-M Li et al ldquoAn improved multi-objective population-based extremal optimization algorithmwith polynomial mutationrdquo Information Sciences vol 330pp 49ndash73 2016

[42] Y-L Hu and L Chen ldquoA nonlinear hybrid wind speedforecasting model using LSTM network hysteretic ELM andDifferential Evolution algorithmrdquo Energy Conversion andManagement vol 173 pp 123ndash142 2018

[43] G-Q Zeng J Chen Y-X Dai L-M Li C-W Zheng andM-R Chen ldquoAn improved Design of fractional order PIDcontroller for automatic regulator voltage system based onmulti-objective extremal optimizationrdquo Neurocomputingvol 160 pp 173ndash184 2015

[44] X Xu R Gu F Dai L Qi and S Wan ldquoMulti-objectivecomputation offloading for internet of vehicles in cloud-edgecomputingrdquo Wireless Networks pp 1ndash19 2019

[45] S Wan X Li Y Xue W Lin and X Xu ldquoEfficient com-putation offloading for internet of vehicles in edge comput-ing-assisted 5G networksrdquo e Journal of Supercomputingpp 1ndash30 2019

[46] B Wang S Wan X Zhang and K-K R Choo ldquoA novelindex for assessing the robustness of integrated electricalnetwork and a natural gas networkrdquo IEEE Access vol 6pp 40400ndash40410 2018

[47] Y Yin F Yu Y Xu L Yu and J Mu ldquoNetwork location-aware service recommendation with random walk in cyber-physical systemsrdquo Sensors vol 17 no 9 p 2059 2017

Complexity 13

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Page 4: Parallel LSTM-Based Regional Integrated Energy System ...downloads.hindawi.com/journals/complexity/2019/7414318.pdf · Characteristics of Regional Integrated Energy System “Information

the load Different energy behaviors correspond to differentload characteristics [26] For example the lighting behaviorin RIES is a typical user energy behavior +rough advancedmeasurement devices the userrsquos lighting behavior can bereflected as a lighting load curve

From the perspective of user energy behavior there aremultiple coupling relationships between multiple energyloads in RIES Taking the userrsquos lighting behavior as anexample if the cloud suddenly covers a certain area andreduces the ambient light intensity it will affect the userrsquoslighting behavior people will increase the lighting equip-ment usage to meet the lighting demand and the lightingload will increase At the same time such cloud movementswill also have an impact on photovoltaic power generationcloud clusters will block the photovoltaic panels which willaffect the irradiance reduction on the photovoltaic panelsand affect the photovoltaic power generation In order toprove the characteristics of this behavioral feature this papertakes the data of three days from January 24 to 26 2016 ofRIES in a certain area of Tianjin for quantitative analysis+ephotovoltaic power output curve and the user lighting loadcurve are shown in Figure 3

+e upper part of Figure 3 is the curve trend of thephotovoltaic power output and the lower part is the curvetrend of the userrsquos lighting load (1) Both the userrsquos lightingload and the photovoltaic power output have a day and nightcycle +e distribution of them is mainly concentrated in thedaytime +e peak of the userrsquos lighting load is generallydistributed between 18 00 and 20 00 and the peak outputof the photovoltaic power supply is generally distributedbetween 12 00 and 14 00 at noon (2) Taking the three-dayshort-term data for comparison analysis the overall outputof photovoltaic power generation on the third day wassignificantly larger than the previous two days Accordinglythe lighting load on the third day between 10 00 and 16 00is reduced compared to the previous two days and it isreduced to the low of the day It shows that when the outputof photovoltaic power generation increases it also affects theuserrsquos energy consumption behavior when the ambient lightintensity is large which means that the userrsquos demand forlighting energy is reduced (3) Take the ultra-short-termdata analysis from 12 00 to 16 00 on the first day andthe photovoltaic power output reaches the peak of one day at13 00 and the lighting load is also reduced to low of thedaytime at this moment During the period from 13 00 to14 00 a sudden downward climb occurs and the lightingload increases at the corresponding time +is ldquomutationrdquophenomenon should be caused by the cloudy weather af-fected by the cloud From the above analysis it can be foundthat there is a strong information coupling correlation be-tween the ldquosourcerdquo photovoltaic power output and theldquocharge endrdquo userrsquos lighting load

33 RIES Multienergy Information Interaction ModeRIES has both horizontal multienergy information couplingcorrelation characteristics and vertical source-load in-formation coupling correlation characteristics From theperspective of information interaction energy management

there are also two horizontal and vertical information in-teraction modes

+e horizontal information interaction mode is based onmultienergy conversion information sharing On the basis ofobtaining the measurement data of the multienergy flowsubsystem the nonlinear artificial intelligence deep learningtechnology is used to process the multienergy data in par-allel and it effectively utilizes the complex shared in-formation of energy conversion identifies the abstractfeatures of the training data and improves the precision ofsubtask prediction of the energy management

+e vertical information interaction mode is based onuser behavior information sharing On the basis of obtainingmulticategory data on both sides of the source and the loadthe deep learning method is used to dig the user-side be-havior characteristics and the associated user behaviorinformation data are taken as input to guide the source-sideenergy management prediction and improve the predictionaccuracy of the energy supply end

4 Multitask Learning Mechanism for DeepLearning Energy Prediction

41 LSTM Recurrent Neural Network

411 Principle of LSTM Cyclic Neural Network +e trainingobjects of the source-charged uncertain resources haveobvious timing characteristics and the sequence data havestrong correlation before and after From the perspective oflearning and training recurrent neural network (RNN) canprocess sequences with temporal correlation of arbitrarylength by using neurons with self-feedback which has ob-vious advantages in processing sequence data [27] +estructure of the RNN is shown in Figure 4(a) +e LSTMnetwork is a variant of RNN that overcomes the traditionalRNN gradient demise problem [28] +e hidden layer is nolonger a simple neural unit but an LSTM unit with a uniquememory mode Each LSTM unit contains a state cell thatdescribes the current state of the LSTM unit [29] c repre-sents the current state quantity of the LSTM unit and h

represents the current output of the LSTM unitLSTM reads and modifies state units by controlling the

forget gates input gates and output gates [30] which are

0

ndash100

250

50

100

150

200

ndash50

ndash200

ndash100

0

100

200

300

P (k

W)

10 20 30 40 50 60 700t (h)

Figure 3 +e relationship between the photovoltaic power curveand the lighting load curve of the user

4 Complexity

typically formed by sigmoid or tanh functions and Hada-mard product operation As shown in Figure 4(b) the forgetgate determines ctminus 1 of the previous moment and how manycomponents of the cell state ct remain at the current timeand the input gate determines how many components of thenetwork input are saved to the cell state ct at the current timext and howmany components of the output gate control cellstate ct are output to the current value ht of the LSTM

+e calculation formula between the variables of theLSTM unit is as follows

ft σ Wfxxi + Wfhhtminus 1 + bf1113872 1113873

it σ Wixxi + Wihhtminus 1 + bi( 1113857

gt ϕ Wgxxt + Wghhtminus 1 + bg1113872 1113873

ct ft ⊙ ctminus 1 + it ⊙gt

ot σ Woxxt + Wohhtminus 1 + bo( 1113857

ht ot ⊙ϕ ct( 1113857

(1)

In the formulaWfxWfhWixWihWgxWghWox andWoh are the weight matrices corresponding to the input of thenetwork activation function bf bi bg and bo are the offsetvectors ⊙ represents the multiplication of the Hadamardproduct matrix σ represents the sigmoid activation functionand ϕ represents the tanh activation function

412 LSTMDeep Learning Training Algorithm +e trainingof LSTM network adopts BP-based backpropagationthrough time (BPTT) [31] +e algorithm steps are as fol-lows (1) Calculate the output value of each neuron forwardFor the LSTMunit the values of the six vectors such asft itgt ct ot and ht are calculated (2) Calculate the error termof each neuron in reverse Similar to traditional RNN thebackpropagation of the LSTM error term consists of twolevels one is at the spatial level and the error term ispropagated to the upper layer of the network the other is atthe time level which propagates back in time For examplefrom the current time of t the error at each moment iscalculated (3) Calculate the gradient of each weightaccording to the corresponding error term and update thenetwork weight parameter (4) Jump to the first step andcontinue to perform the first 3 steps until the network erroris less than the given value

42 LSTM Network Multitask Learning MechanismMultitask learning (MTL) as an inductive migrationmechanism avoids the underfitting of model training byusing the correlation between multiple tasks digging thehidden common features and improving the generalizationlearning ability of the model [32] RIES source wind andlight prediction and load-end electricity gas heat and coldload prediction can be regarded as different tasks in mul-titask learning [33] Multiple tasks share some common dataor model structures that is there is a certain correlationbetween the multienergy prediction tasks +e multitasklearning prediction model based on the LSTM networkproposed in this paper shares the hidden layer nodes amongall prediction tasks through the hard parameter sharingmethod and trains the training data corresponding to themultienergy prediction tasks at the same time consideringthe correlation information between tasks to improve thelearning ability of all predictive tasks [34] A comparison ofsingle-task and multitask learning based on the LSTMnetwork is shown in Figure 5 +e multisource historicaldata in the figure is a collection of five types of historical dataincluding wind power photovoltaic electric load heat loadand cold load

5 Parallel LSTM-Based RIES InformationInteractive Energy Prediction Method

51 RIES Multienergy Prediction Model Design RIES hasuncertainties at both the source and the load end so accurateprediction of uncertain resources is the basis for imple-menting RIES energy optimization management andscheduling [35] +e source mainly includes wind powerprediction and photovoltaic power prediction and the loadmainly includes electric load prediction thermal load pre-diction and cold load prediction

As described in the second section of this paper there areintricate mutual coupling relationships between multipleenergy loads of RIES and the prediction tasks of uncertainresources on both sides of the source and load have strongcorrelation It can effectively improve the overall predictionaccuracy of the RIES uncertain resources by learning thiscoupling relationship through joint prediction training [36]From the perspective of RIES multienergy-load informationinteraction this paper proposes the concept of ldquoparallel

xtndash1 xt xt+1

htndash2

ctndash2

hh h

cc c

htndash1 ht ht+1

ctndash1 ct ct+1

(a)

tanh

tanh

xt

ht

ctctndash1

Forgetgate

ft it ot

ct

htndash1

Inputgate

Outputgate

σ σ σ

(b)

Figure 4 Schematic diagram of RNN and LSTM (a) +e structure of the RNN (b) +e structure of the LSTM cell

Complexity 5

predictionrdquo and utilizes the strong nonlinear learning abilityand historical memory advantages of LSTM deep learningnetwork to conduct integrated prediction of uncertain re-sources at both source and load As shown in Figure 6 thehorizontal information interaction based on multienergyconversion information sharingmainly utilizes themultitasklearning parameter sharing mechanism of the parallel LSTMprediction network the vertical information interactionbased on the user behavior information sharing mainly usesthe user behavior data to guide source-load multitask pre-diction as an additional unified input information Para-metric sharing is weight sharing which takes advantage ofthe correlation between subtasks Input sharing refers todata sharing Different subtasks share historical power datasuch as wind power photovoltaic power power loadthermal load and cooling load

52 Data Preprocessing +e training samples select the windpower generation time series on the source side the photo-voltaic power generation time series the charge side on the loadside the heat load the cold load time series and the illumi-nation load series data Considering the input and output rangeof the nonlinear activation function in the model in order toavoid neuron saturation the data are normalized and mappedbetween [minus 1 1] as shown in the following equation

xprime x minus xmax + xmin( 11138572

xmax minus xmin( 11138572 (2)

In the equation x is the actual input or output data andxmax and xmin are the maximum and minimum values of thevariable respectively

53 Task Construction +e essence of using the multitasklearning method for multienergy-load integration predictionlies in the close correlation of each prediction subtask [37]+e traditional single-task prediction method uses a singledimension of historical observations to predict its futureoutput and the model output is a single attribute +emultitask learning integration prediction uses the multidi-mensional data on both sides of the source and the load as a

common input to effectively utilize the sample On the otherhand through the shared hidden layer node divisionmethodthe five attributes of wind photovoltaic power generationelectricity heat and cold load are simultaneously predictedto play the role of inductive bias and improve predictionaccuracy

54 Model Training +e LSTM-based source-charge in-tegrated prediction model mainly needs to determine thefollowing parameters input layer dimension input layertime step hidden layer number each hidden layer di-mension and output layer dimension [38]

+e time step of the input layer is generally the length ofthe variable time series used for the source-load integrationprediction According to the periodic characteristics of theprediction object sequence the input layer time step is set to24 that is the historical data whose resolution ratio is 1 hduring the first seven days before the input is used forforecasting and it takes up one input node per day+e inputlayer dimension is the number of nodes +e input layernode of this paper contains six objects wind power gen-eration time series photovoltaic power generation timeseries charge side load heat load cold load time series andlighting load series which constitutes totally 42 nodes so theinput layer dimension is set to 42 Since this paper simul-taneously predicts six attributes of wind power photovol-taic electricity heat and cold load the output layer has 5nodes that is the output layer dimension is set to 5 In thispaper based on experience and multiple trial and com-parison the number of hidden layers is set to 2 and thedimension of the hidden layer is set to 30 +e selection ofthe activation function is as shown in Figure 4(a) thebatch_size is set to 42 the numbers of iterations of an epochis set to 8784 the learning rate is set to 080 and the learningdelay rate is set to 005 +e network training function usesthe BPTT algorithm mentioned above

55 Evaluation Indicators Multienergy-load integrationforecasting simultaneously trains multiple forecastingtasks To comprehensively measure the forecasting effect itis necessary to evaluate the integrated forecasting accuracyas a whole [39] In this paper the average accuracyevaluation index based on weight coefficient is proposedFor the importance degree of different energy on the twosides of the RIES the corresponding weight coefficient isgiven to different energy output types For an RIES theuncertainties of the source-side wind light and otheruncertain resources are large and have a great influence onthe stable operation of the system and the energy dispatchmanagement +erefore the source-side wind forecastingis set to a higher weight in addition as the power gridplays a leading role in the integrated energy system byvirtue of its perfect architecture and central hub advan-tages correspondingly the power load forecast is set to ahigher weight

+is paper proposes three calculation indicators namelymean absolute percentage error (MAPE) mean accuracy

x1 x2 xm

x1 x2 xn

x1 x2 xk

Hiddenlayer

Multitask learningOutputlayer

Inputlayer

Single-task learning

Wind powerhistorical data

Electrical loadhistorical data

Multisourcehistorical data

Figure 5 Comparison of single-task and multitasking learningbased on the LSTM network

6 Complexity

(MA) and weightedmean accuracy (WMA) as shown in thefollowing equation

MAPE 1n

1113944

n

i1

P(i) minus 1113954P(i)

P(i)

11138681113868111386811138681113868111386811138681113868

11138681113868111386811138681113868111386811138681113868times 100

MA 1 minus MAPE

WMA λ1MA1 + λ2MA2 + middot middot middot + λkMAk

(3)

In the equation P(i) and 1113954P(i) respectively representthe actual energy value and the predicted energy value at thefirst moment n is the sample data amount MAk representsthe prediction accuracy of the k prediction task and λk

represents the weight coefficient of the k prediction task Inthis paper the weight ratios of the energy predictions onboth sides of the source and the load are set as shown inTable 1

6 Simulation Analysis

61 Data Description +e experimental data in this paperare derived from the measured data provided by RIES in acertain area of Tianjin +e RIES consists of a power systema thermal system and a natural gas system +e source sideincludes energy supply equipment such as wind turbinesand photovoltaic power generation equipment Couplinglinks include energy conversion equipment such as cold andthermal power supply equipment electric boilers and gasboilers Energy demand mainly includes power load heatload and cooling load +e load side and the energy storageside contain devices such as electrical energy storage andthermal energy storage to improve the flexibility of thesystem In addition in order to prove the validity of thesource-load vertical information interaction the historicaldata of the lighting load reflecting the typical electricity

usage behavior of the user are also sampled All historicalenergy data sampling interval is 1 h data from July 2015 toJune 2016 are used as training set data and data from July2016 are used as test set data with 1 h as time step based onthe historical data of the first 7 days the energy supply andenergy demand on both sides of the source and the load inthe next day are predicted +e total 57168 samples weredivided into two parts (92 as the train set and 8 as the testset) +e forecasting models were trained by the train datasetand validated on the test dataset +e statistical informationof the above dataset is shown in Table 2

62 Analysis of Examples +is section analyzes the effect ofRIES multienergy information interactive energy prediction+e selected sunny day and rainy day of the test set wereanalyzed separately +e energy prediction results of windpower output photovoltaic output electric load heat loadand cold load on both sides of the source and the load in thetwo scenarios are shown in Figures 7 and 8 respectively

+e prediction results of Figures 7 and 8 show that theproposed source-load parallel LSTM learning model canobtain ideal results in wind energy photovoltaic electricload thermal load and cold load energy prediction in sunnyday scenario and rainy day scenario which confirms theeffectiveness of the model proposed in this paper +eprediction effect of each subtask in the parallel predictionresults shows a large difference Compared with the windand light prediction on the source side the predicted curvesof the charge heat and cold load on the charge side arecloser to the actual sequence curve which is due to theintermittent and random nature of source-side wind powerand photovoltaics

Comparing Figures 7(c)ndash7(e) it can be found that theelectric load is more volatile than the thermal load and thecold load which is because the thermal inertia and cold

LSTM 1

LSTMn

Forecast 1input data

Forecast ninput data

Parameters sharing

Forecast 1output data

Forecast noutput data

Inputssharing

Use

r beh

avio

r

Mul

tiene

rgy

sour

ce-lo

ad in

tegr

ated

fore

casti

ng re

sults

Figure 6 Interactive energy prediction model of RIES information based on parallel LSTM

Complexity 7

inertia of the thermal system are stronger [39] and thevolatility is also even smaller +ermal inertia means that thedynamic change of the thermal system is a slow process froma time scale Compared with the power system the heatenergy transmission exhibits a delay effect and the upperand lower fluctuations of the heat load exhibit an inertialeffect and the cold inertia is similar Because of the existenceof thermal inertia and cold inertia the thermal load in thethermal system is less random in the short-term predicted

time scale showing a characteristic of smooth variation soits prediction uncertainty is also small and the thermal loadand the cold load are relative to the electricity +e pre-diction error of the load is also relatively small ComparingFigures 7 and 8 it can be found that the uncertainty of actualphotovoltaic power and actual wind power in rainy dayscenes is larger than that of sunny days+e forecast curve ofphotovoltaic and wind power in sunny scenes is closer to thetrue value the electric load and heat load in rainy scenes

Table 1 Weight settings of source-load double-sided energy prediction tasks

Prediction task Wind power PV power Electric load Heat load Cold loadWeight coefficient 025 025 02 015 015

Table 2 +e description of simulation dataset

Data type Max (kW) Median (kW) Min (kW) Mean (kW) Std (kW) Train set size Test set sizeWind power 317561 233774 36612 216102 78964 8784 744Photovoltaic power 235000 0000 0000 42507 70557 8784 744Electric load 696585 584923 209872 523803 139174 8784 744Heat load 148473 92719 60532 100076 27240 8784 744Cold load 436964 223213 39046 221949 135304 8784 744Lighting load 03333 50000 2778 44792 27734 8784 744Multisource fusion mdash mdash mdash mdash mdash 52704 4464

True valuePredictive value

50

100

150

200

250

300

Win

d po

wer

(kW

)

5 10 15 20 250t (h)

(a)

True valuePredictive value

5 10 15 20 250t (h)

0

50

100

150

200

250

PV p

ower

(kW

)

(b)

True valuePredictive value

5 10 15 20 250t (h)

200

300

400

500

600

700

Elec

tric

load

(kW

)

(c)

True valuePredictive value

5 10 15 20 250t (h)

40

60

80

100

120

140

Hea

d lo

ad (k

W)

(d)

True valuePredictive value

5 10 15 20 250t (h)

0

100

200

300

400

500

Col

d lo

ad (k

W)

(e)

Figure 7 +e results of source-load bilateral energy prediction within a sunny day (a) Prediction of wind power (b) Prediction of PVpower (c) Prediction of electric load (d) Prediction of heat load (e) Prediction of cold load

8 Complexity

should increase as a whole +is is due to the cold weather inthe rainy scenes which affects the userrsquos energy use behaviorand increases the userrsquos demand for electricity and heat thusincreasing the electrical load and heat load

621 Comparative Analysis of Multienergy Load InformationInteraction In order to verify the effect of interactivelearning of horizontal and vertical multienergy informationthe parallel LSTM predictions considering user energy be-havior and parallel LSTM prediction without consideringuser energy behavior and the results of single LSTM pre-diction without considering user energy behavior are pro-posed to conduct a comparative analysis In order to ensurethe fairness of the experimental test this paper uses the samedataset for training and testing

+e comparison experiment setup is shown in Table 3 Inthe parallel LSTM learning without considering the userrsquosenergy behavior the input sample is based on the originalprediction model A and the lighting load data reflecting theuser behavior in the input sample are not considered +eprediction is based only on the historical data of the pre-dicted object the model C does not consider the singleLSTMprediction of the userrsquos energy behavior and the inputand output of each subprediction model only consider theinput and output of a single energy and there is also noparameter sharing between the LSTM unit of the different

submodels Model B uses parallel LSTM learning +ere isparameter sharing among each subtask which belongs to amultitask learning Model C uses a single LSTM learningand each subtask does not affect each other It belongs to asingle-task learning

+e prediction accuracy of the multienergy charge in-formation interaction comparison experiment is shown inTable 4 and Figure 9 +rough the analysis of Table 4 andFigure 9 it shows that the average prediction accuracy of theweights of the models A B and C are 09009 08917 and08728 respectively +e prediction result of the model A isthe best the model B is the second and the model C is thethird +at is considering the interaction of vertical andhorizontal information the energy predictions on both sidesof the source and the load show better prediction accuracy

Comparing the prediction results of Model B and ModelC for the five subprediction tasks when the different qualityenergies are predicted together in parallel a higher precisionthan the single prediction is obtained It can be seen thatparallel LSTM learning considering horizontal multienergyinformation interaction utilizes the coupled data betweendifferent systems to fully exploit the interactive informationbetween different quality energy sources of the integratedenergy system so that the prediction performance of acertain energy can be improved In addition considering themultienergy horizontal information interaction the pre-dicted accuracy of the load-side electrical load heat load and

True valuePredictive value

5 10 15 20 250t (h)

100

150

200

250

300

350W

ind

pow

er (k

W)

(a)

True valuePredictive value

5 10 15 20 250t (h)

0

20

40

60

80

100

120

PV p

ower

(kW

)(b)

True valuePredictive value

5 10 15 20 250t (h)

300

400

500

600

700

Elec

tric

load

(kW

)

(c)

True valuePredictive value

5 10 15 20 250t (h)

60

80

100

120

140

160

Hea

t loa

d (k

W)

(d)

True valuePredictive value

5 10 15 20 250t (h)

0

100

200

300

400

500

Col

d lo

ad (k

W)

(e)

Figure 8+e results of source-load bilateral energy prediction within a rainy day (a) Prediction of wind power (b) Prediction of PV power(c) Prediction of electric load (d) Prediction of heat load (e) Prediction of cold load

Complexity 9

cold load are increased by 313 253 and 267 re-spectively while the predicted accuracy of the source-sidewind power and photovoltaic is increased by 152 and146 and the energy prediction accuracy of the charge sideis relatively high It can be seen that the source-side un-certainty resources are also affected by other energy in-teractions in the system but the impact is small while thecharge side demand is largely affected by other energy in-teractions in the system And the information interaction ismore intimate between each other

Comparing the prediction results of Model A and ModelB when the userrsquos energy behavior information interactsbetween the source and the load in vertical direction theaverage weight accuracy of the overall prediction is increasedby 103 Comparing the performance improvement effectsof the five subprediction tasks it can be found that thepercentage improvement of PV prediction accuracy is185 and the improvement effect is relatively best fol-lowed by the electrical load while the prediction accuracy ofwind power heat load and cooling load is not obvious Itcan be seen that considering the source-load vertical in-formation interaction can improve the energy predictionaccuracy on both sides of the source and the load but thespecific user energy behavior has different differences in theenergy prediction of different types of users and the user

lighting behavior is obviously related to photovoltaic powergeneration and user electricity +e interaction betweenloads is more tight and its effect on PV prediction andelectric load forecasting is more obvious

622 Comparison of Deep Neural Network Algorithms withOther Algorithms In order to verify the prediction effect ofthe LSTM deep neural network algorithm the LSTM modeland the autoregressive integrated moving average model(ARIMA) algorithm and the classic BP neural network(BPNN) algorithm in this paper are proposed +e simu-lation results of them were compared and analyzed In orderto ensure the fairness of the experimental test this paper usesthe same dataset for training and testing and both use asingle-task prediction strategy +e prediction results of thethree algorithms are shown in Table 5 and Figure 10

Observing the prediction results the prediction effect ofARIMA algorithm is the most general and the predictionaccuracy of BPNN algorithm is much higher than that ofARIMA algorithm +e prediction accuracy of each task ofLSTM deep learning prediction algorithm proposed in thispaper is the highest +e ARIMA algorithm can not effec-tively use the historical information in the long-term se-quence because it only uses a part of the historicalinformation to train which leads to the large predictionerror +e LSTM algorithm can fully utilize its strongnonlinear learning ability and long-term and short-termmemory advantages to exploit the information value of long-term historical data so the grasp of the energy prediction lawis more precise

In addition comparing the energy prediction results ofthe LSTM algorithm and the BPNN algorithm one by onethe LSTM algorithm predicts that the accuracy of thethermal load and the cold load prediction is more obvious Itcan be seen that due to the greater inertia of the thermalsystem the time scale of the change law is longer whichmakes the prediction accuracy of the long-term and short-term deep learning on the heat load and the cold loadsignificantly higher than the electric load

In order to evaluate the computational efficiency ofalgorithms four forecasting scenarios are performed tentimes so that the forecasting results can reliably be sum-marized as in Table 6 Parallel LSTM and Single LSTM havelonger training time but the prediction time is kept within

Table 3 Contrast experiment setting of multisource source information interaction

Model Vertical interaction Horizontal interactionModel A User energy behavior is considered Parallel LSTM learningModel B User energy behavior is not considered Parallel LSTM learningModel C User energy behavior is not considered Single-LSTM learning

Table 4 +ree kinds of model prediction results of information interaction comparison experiment

Precision Wind power PV power Electric load Heat load Cold load Mean precisionModel A 08602 08735 09236 09445 09407 09009Model B 08542 08576 09113 09401 09368 08917Model C 08414 08453 08836 09169 09124 08728

Model AModel BModel C

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

PV Electric Heat Cold MeanWindPrediction task

Figure 9 Comparison of prediction effects of three models

10 Complexity

one minute and can meet the requirements of short-termforecasting Further analysis shows that the parallel trainingtime and prediction time of parallel LSTM is greatly im-proved compared to single LSTM due to the combination offive subtasks into one multitask learning In summary al-though the LSTM algorithm is time consuming in trainingtime for short-term prediction the training process doesnot require online completion and the prediction time ofthe LSTM algorithm especially parallel LSTM can fullymeet the requirements of short-term prediction Consid-ering the prediction accuracy the sacrifice of the LSTMalgorithm in computing time is worthwhile

7 Conclusion

In this paper the horizontal and vertical information interactioncharacteristics between multiple energy loads in RIES areconsidered Using LSTM network deep learning algorithm a

parallel LSTM-based multienergy-load information interactiveshort-term energy prediction method is proposed+rough thesimulation of the test data the following conclusions are drawn

(1) +e ldquoinformation interactionrdquo strategy proposed inthis paper can effectively improve the predictionperformance of source-load energy prediction inregional integrated energy systems

(2) Parallel prediction considering horizontal in-teraction mainly affects the prediction effect on theload side while the information sharing of user il-lumination behavior considering vertical interactionmainly affects the improvement of PV predictioneffect

(3) Compared with the conventional prediction algo-rithm the LSTM deep learning algorithm can learnthe deep hidden information of historical dataimprove the prediction accuracy while meetingcalculation time and prove the effectiveness of thealgorithm in the prediction of time series featuresequence

However the performance of the proposed parallelLSTM algorithm may be further enhanced by some evo-lutionary algorithms or multiobjective optimization algo-rithms such as differential evolution algorithm [40]multiobjective extremal optimization [41] and evolutionarymultiobjective optimization algorithms [42ndash47] Of coursesome significant topics are worth researching on the opti-mization of the proposed parallel LSTM by multiobjectiveoptimization

Data Availability

+e data used to support the findings of this study areavailable from the corresponding author upon request

Conflicts of Interest

+e authors declare no potential conflicts of interest

Acknowledgments

+is work was fully supported by the National Key Researchand Development Program of China (no 2018YFB0904200)

References

[1] A Q Huang M L Crow G T Heydt J P Zheng andS J Dale ldquo+e future renewable electric energy delivery andmanagement (FREEDM) system the energy internetrdquo Pro-ceedings of the IEEE vol 99 no 1 pp 133ndash148 2011

[2] I Colak S Sagiroglu G Fulli M Yesilbudak and C CovrigldquoA survey on the critical issues in smart grid technologiesrdquo

Table 5 Comparison of prediction results of three algorithms

Precision Wind power PV power Electric load Heat load Cold load Mean precisionLSTM 08414 08453 08836 09169 09124 08728BPNN 08216 08314 08695 08912 08889 08542ARMIA 07733 07824 08282 08454 08408 08075

Prediction precision ofwind powerPrediction precision ofPV powerPrediction precision ofelectric load

Prediction precision ofhead loadPrediction precision ofcold loadWeight average predictionprecision

07

075

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

BPNN ARMIALSTMPrediction algorithm

Figure 10 Comparison of prediction effects of three algorithms

Table 6 Comparison of computational efficiency of fouralgorithms

Algorithm Training timeaverage (sec)

Prediction timeaverage (sec)

Parallel LSTM 436374 1002Single LSTM 1218337 3662BPNN 1094911 2253ARMIA 54696 215

Complexity 11

Renewable and Sustainable Energy Reviews vol 54 pp 396ndash405 2016

[3] C Tu X He Z Shuai and F Jiang ldquoBig data issues in smartgrid-a reviewrdquo Renewable and Sustainable Energy Reviewsvol 79 pp 1099ndash1107 2017

[4] Y Zhang R Yang K Zhang H Jiang and J J ZhangldquoConsumption behavior analytics-aided energy forecastingand dispatchrdquo IEEE Intelligent Systems vol 32 no 4pp 59ndash63 2017

[5] J Deng F Wang Y Chen and X Zhao ldquoFrom industry 40 toenergy 50 the concept connotation and system framework ofintelligent energy systemsrdquo Journal of Automation vol 41no 12 pp 2003ndash2016 2015

[6] Q Sun L Shi Y Ni D Si and J Zhu ldquoAn enhanced cas-cading failure model integrating data mining techniquerdquoProtection and Control of Modern Power Systems vol 2 2017

[7] D Zheng A T Eseye J Zhang and H Li ldquoShort-term windpower forecasting using a double-stage hierarchical ANFISapproach for energy management in microgridsrdquo Protectionand Control of Modern Power Systems vol 2 pp 1ndash10 2017

[8] Z Li L Ye Y Zhao X Song J Teng and J Jin ldquoShort-termwind power prediction based on extreme learning machinewith error correctionrdquo Protection and Control of ModernPower Systems vol 1 no 1 pp 1ndash8 2016

[9] H Z Wang Z X Lei X Zhang B Zhou and J C Peng ldquoAreview of deep learning for renewable energy forecastingrdquoEnergy Conversion and Management vol 198 pp 1ndash16 2019

[10] H Z Wang G B Wang G Q Li J C Peng and Y T LiuldquoDeep belief network based deterministic and probabilisticwind speed forecasting approachrdquo Applied Energy vol 182pp 80ndash93 2016

[11] M-R Chen G-Q Zeng K-D Lu and J Weng ldquoA two-layernonlinear combination method for short-term wind speedprediction based on ELM ENN and LSTMrdquo IEEE Internet ofings Journal vol 6 no 4 pp 6997ndash7010 2019

[12] H Liu X Mi and Y Li ldquoSmart multi-step deep learningmodel for wind speed forecasting based on variational modedecomposition singular spectrum analysis LSTM networkand ELMrdquo Energy Conversion and Management vol 159pp 54ndash64 2018

[13] J Chen G-Q Zeng W Zhou W Du and K-D Lu ldquoWindspeed forecasting using nonlinear-learning ensemble of deeplearning time series prediction and extremal optimizationrdquoEnergy Conversion and Management vol 165 pp 681ndash6952018

[14] W Kong Z Y Dong Y Jia D J Hill Y Xu and Y ZhangldquoShort-term residential load forecasting based on LSTM re-current neural networkrdquo IEEE Transactions on Smart Gridvol 10 no 1 2017

[15] H Gao W Huang X Yang Y Duan and Y Yin ldquoTowardservice selection for workflow reconfiguration an interface-based computing solutionrdquo Future Generation ComputerSystems vol 87 pp 298ndash311 2018

[16] L Qi Q He F Chen et al ldquoFinding all you need web APIsrecommendation in web of things through keywords searchrdquoIEEE Transactions on Computational Social Systems vol 6no 5 2019

[17] L Qi R Wang C Hu S Li Q He and X Xu ldquoTime-awaredistributed service recommendation with privacy-preserva-tionrdquo Information Sciences vol 480 pp 354ndash364 2019

[18] Y Zhang and Q Yang ldquoAn overview of multi-task learningrdquoNational Science Review vol 5 no 1 pp 34ndash47 2018

[19] A Argyriou M T Evgeniou and M Pontil ldquoConvex multi-task feature learningrdquo Machine Learning vol 73 no 3pp 243ndash272 2008

[20] T Jebara ldquoMultitask sparsity via maximum entropy dis-criminationrdquo Journal of Machine Learning Research vol 12no 1 pp 75ndash110 2011

[21] L Qi Y Chen Y Yuan S Fu X Zhang and X Xu ldquoA QoS-aware virtual machine scheduling method for energy con-servation in cloud-based cyber-physical systemsrdquo WorldWide Web vol 5 pp 1ndash23 2019

[22] R Zhang P Xie C Wang G Liu and S Wan ldquoClassifyingtransportation mode and speed from trajectory data via deepmulti-scale learningrdquo Computer Networks vol 162 p 1068612019

[23] W Wang D Wang H Jia et al ldquoReview of steady-stateanalysis of typical regional integrated energy system under thebackground of energy internetrdquo Proceedings of the CSEE12pp 3292ndash3306 2016

[24] S Ding S Qu Y Xi and S Wan ldquoStimulus-driven andconcept-driven analysis for image caption generationrdquoNeurocomputing 2019

[25] P Vrba V Marik P Siano et al ldquoA review of agent andservice-oriented concepts applied to intelligent energy sys-temsrdquo IEEE Transactions on Industrial Informatics vol 10no 3 pp 1890ndash1903 2014

[26] W Jia C Kang and Q Chen ldquoAnalysis on demand-sideinteractive response capability for power system dispatch in asmart grid frameworkrdquo Electric Power Systems Researchvol 90 pp 11ndash17 2012

[27] Z C Lipton J Berkowitz and C Elkan ldquoA critical review ofrecurrent neural networks for sequence learningrdquo CoRR2015 httpsarxivorgpdf150600019pdf

[28] T N Sainath O Vinyals A Senior and H Sak ldquoCon-volutional long short-term memory fully connected deepneural networksrdquo in Proceedings of the 2015 IEEE In-ternational Conference on Acoustics Speech and Signal Pro-cessing (ICASSP) pp 4580ndash4584 IEEE Brisbane AustraliaApril 2015

[29] Y Guo Z Cheng L Nie YWang J Ma andM KankanhallildquoAttentive long short-term preference modeling for person-alized product searchrdquo ACM Transactions on InformationSystems (TOIS) vol 37 no 2 p 19 2019

[30] A Graves ldquoLong short-term memoryrdquo Neural Computationvol 9 pp 1735ndash1780 1997

[31] F A Gers J Schmidhuber and F Cummins ldquoLearning toforget continual prediction with LSTMrdquo Neural Computa-tion vol 12 no 10 pp 2451ndash2471 2000

[32] M L Seltzer and J Droppo ldquoMulti-task learning in deepneural networks for improved phoneme recognitionrdquo inProceedings of the 2013 IEEE International Conference onAcoustics Speech and Signal Processing pp 6965ndash6969 IEEEVancouver Canada May 2013

[33] S Wan Y Zhao T Wang Z Gu Q H Abbasi andK-K R Choo ldquoMulti-dimensional data indexing and rangequery processing via Voronoi diagram for internet of thingsrdquoFuture Generation Computer Systems vol 91 pp 382ndash3912019

[34] Z Gao D Y Wang S H Wan H Zhang and Y L WangldquoCognitive-inspired class-statistic matching with triple-con-strain for camera free 3D object retrievalrdquo Future GenerationComputer Systems vol 94 pp 641ndash653 2019

[35] C A Kang A R Brandt and L J Durlofsky ldquoOptimaloperation of an integrated energy system including fossil fuel

12 Complexity

power generation CO2 capture and windrdquo Energy vol 36no 12 pp 6806ndash6820 2011

[36] Q Zhang SWan BWang DW Gao and HMa ldquoAnomalydetection based on randommatrix theory for industrial powersystemsrdquo Journal of Systems Architecture vol 95 pp 67ndash742019

[37] Z Gao H Z Xuan H Zhang S Wan and K K R ChooldquoAdaptive fusion and category-level dictionary learningmodel for multi-view human action recognitionrdquo IEEE In-ternet of ings Journal p 1 2019

[38] WWu K Chen Y Qiao and Z Lu ldquoProbabilistic short-termwind power forecasting based on deep neural networksrdquo inProceedings of the 2016 International Conference on Proba-bilistic Methods Applied to Power Systems (PMAPS) pp 1ndash8Beijing China October 2016

[39] K M Powell A Sriprasad W J Cole and T F EdgarldquoHeating cooling and electrical load forecasting for a large-scale district energy systemrdquo Energy vol 74 pp 877ndash8852014

[40] A Peimankar S J Weddell T Jalal and A C LapthornldquoMulti-objective ensemble forecasting with an application topower transformersrdquo Applied Soft Computing vol 68pp 233ndash248 2018

[41] G-Q Zeng J Chen L-M Li et al ldquoAn improved multi-objective population-based extremal optimization algorithmwith polynomial mutationrdquo Information Sciences vol 330pp 49ndash73 2016

[42] Y-L Hu and L Chen ldquoA nonlinear hybrid wind speedforecasting model using LSTM network hysteretic ELM andDifferential Evolution algorithmrdquo Energy Conversion andManagement vol 173 pp 123ndash142 2018

[43] G-Q Zeng J Chen Y-X Dai L-M Li C-W Zheng andM-R Chen ldquoAn improved Design of fractional order PIDcontroller for automatic regulator voltage system based onmulti-objective extremal optimizationrdquo Neurocomputingvol 160 pp 173ndash184 2015

[44] X Xu R Gu F Dai L Qi and S Wan ldquoMulti-objectivecomputation offloading for internet of vehicles in cloud-edgecomputingrdquo Wireless Networks pp 1ndash19 2019

[45] S Wan X Li Y Xue W Lin and X Xu ldquoEfficient com-putation offloading for internet of vehicles in edge comput-ing-assisted 5G networksrdquo e Journal of Supercomputingpp 1ndash30 2019

[46] B Wang S Wan X Zhang and K-K R Choo ldquoA novelindex for assessing the robustness of integrated electricalnetwork and a natural gas networkrdquo IEEE Access vol 6pp 40400ndash40410 2018

[47] Y Yin F Yu Y Xu L Yu and J Mu ldquoNetwork location-aware service recommendation with random walk in cyber-physical systemsrdquo Sensors vol 17 no 9 p 2059 2017

Complexity 13

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Page 5: Parallel LSTM-Based Regional Integrated Energy System ...downloads.hindawi.com/journals/complexity/2019/7414318.pdf · Characteristics of Regional Integrated Energy System “Information

typically formed by sigmoid or tanh functions and Hada-mard product operation As shown in Figure 4(b) the forgetgate determines ctminus 1 of the previous moment and how manycomponents of the cell state ct remain at the current timeand the input gate determines how many components of thenetwork input are saved to the cell state ct at the current timext and howmany components of the output gate control cellstate ct are output to the current value ht of the LSTM

+e calculation formula between the variables of theLSTM unit is as follows

ft σ Wfxxi + Wfhhtminus 1 + bf1113872 1113873

it σ Wixxi + Wihhtminus 1 + bi( 1113857

gt ϕ Wgxxt + Wghhtminus 1 + bg1113872 1113873

ct ft ⊙ ctminus 1 + it ⊙gt

ot σ Woxxt + Wohhtminus 1 + bo( 1113857

ht ot ⊙ϕ ct( 1113857

(1)

In the formulaWfxWfhWixWihWgxWghWox andWoh are the weight matrices corresponding to the input of thenetwork activation function bf bi bg and bo are the offsetvectors ⊙ represents the multiplication of the Hadamardproduct matrix σ represents the sigmoid activation functionand ϕ represents the tanh activation function

412 LSTMDeep Learning Training Algorithm +e trainingof LSTM network adopts BP-based backpropagationthrough time (BPTT) [31] +e algorithm steps are as fol-lows (1) Calculate the output value of each neuron forwardFor the LSTMunit the values of the six vectors such asft itgt ct ot and ht are calculated (2) Calculate the error termof each neuron in reverse Similar to traditional RNN thebackpropagation of the LSTM error term consists of twolevels one is at the spatial level and the error term ispropagated to the upper layer of the network the other is atthe time level which propagates back in time For examplefrom the current time of t the error at each moment iscalculated (3) Calculate the gradient of each weightaccording to the corresponding error term and update thenetwork weight parameter (4) Jump to the first step andcontinue to perform the first 3 steps until the network erroris less than the given value

42 LSTM Network Multitask Learning MechanismMultitask learning (MTL) as an inductive migrationmechanism avoids the underfitting of model training byusing the correlation between multiple tasks digging thehidden common features and improving the generalizationlearning ability of the model [32] RIES source wind andlight prediction and load-end electricity gas heat and coldload prediction can be regarded as different tasks in mul-titask learning [33] Multiple tasks share some common dataor model structures that is there is a certain correlationbetween the multienergy prediction tasks +e multitasklearning prediction model based on the LSTM networkproposed in this paper shares the hidden layer nodes amongall prediction tasks through the hard parameter sharingmethod and trains the training data corresponding to themultienergy prediction tasks at the same time consideringthe correlation information between tasks to improve thelearning ability of all predictive tasks [34] A comparison ofsingle-task and multitask learning based on the LSTMnetwork is shown in Figure 5 +e multisource historicaldata in the figure is a collection of five types of historical dataincluding wind power photovoltaic electric load heat loadand cold load

5 Parallel LSTM-Based RIES InformationInteractive Energy Prediction Method

51 RIES Multienergy Prediction Model Design RIES hasuncertainties at both the source and the load end so accurateprediction of uncertain resources is the basis for imple-menting RIES energy optimization management andscheduling [35] +e source mainly includes wind powerprediction and photovoltaic power prediction and the loadmainly includes electric load prediction thermal load pre-diction and cold load prediction

As described in the second section of this paper there areintricate mutual coupling relationships between multipleenergy loads of RIES and the prediction tasks of uncertainresources on both sides of the source and load have strongcorrelation It can effectively improve the overall predictionaccuracy of the RIES uncertain resources by learning thiscoupling relationship through joint prediction training [36]From the perspective of RIES multienergy-load informationinteraction this paper proposes the concept of ldquoparallel

xtndash1 xt xt+1

htndash2

ctndash2

hh h

cc c

htndash1 ht ht+1

ctndash1 ct ct+1

(a)

tanh

tanh

xt

ht

ctctndash1

Forgetgate

ft it ot

ct

htndash1

Inputgate

Outputgate

σ σ σ

(b)

Figure 4 Schematic diagram of RNN and LSTM (a) +e structure of the RNN (b) +e structure of the LSTM cell

Complexity 5

predictionrdquo and utilizes the strong nonlinear learning abilityand historical memory advantages of LSTM deep learningnetwork to conduct integrated prediction of uncertain re-sources at both source and load As shown in Figure 6 thehorizontal information interaction based on multienergyconversion information sharingmainly utilizes themultitasklearning parameter sharing mechanism of the parallel LSTMprediction network the vertical information interactionbased on the user behavior information sharing mainly usesthe user behavior data to guide source-load multitask pre-diction as an additional unified input information Para-metric sharing is weight sharing which takes advantage ofthe correlation between subtasks Input sharing refers todata sharing Different subtasks share historical power datasuch as wind power photovoltaic power power loadthermal load and cooling load

52 Data Preprocessing +e training samples select the windpower generation time series on the source side the photo-voltaic power generation time series the charge side on the loadside the heat load the cold load time series and the illumi-nation load series data Considering the input and output rangeof the nonlinear activation function in the model in order toavoid neuron saturation the data are normalized and mappedbetween [minus 1 1] as shown in the following equation

xprime x minus xmax + xmin( 11138572

xmax minus xmin( 11138572 (2)

In the equation x is the actual input or output data andxmax and xmin are the maximum and minimum values of thevariable respectively

53 Task Construction +e essence of using the multitasklearning method for multienergy-load integration predictionlies in the close correlation of each prediction subtask [37]+e traditional single-task prediction method uses a singledimension of historical observations to predict its futureoutput and the model output is a single attribute +emultitask learning integration prediction uses the multidi-mensional data on both sides of the source and the load as a

common input to effectively utilize the sample On the otherhand through the shared hidden layer node divisionmethodthe five attributes of wind photovoltaic power generationelectricity heat and cold load are simultaneously predictedto play the role of inductive bias and improve predictionaccuracy

54 Model Training +e LSTM-based source-charge in-tegrated prediction model mainly needs to determine thefollowing parameters input layer dimension input layertime step hidden layer number each hidden layer di-mension and output layer dimension [38]

+e time step of the input layer is generally the length ofthe variable time series used for the source-load integrationprediction According to the periodic characteristics of theprediction object sequence the input layer time step is set to24 that is the historical data whose resolution ratio is 1 hduring the first seven days before the input is used forforecasting and it takes up one input node per day+e inputlayer dimension is the number of nodes +e input layernode of this paper contains six objects wind power gen-eration time series photovoltaic power generation timeseries charge side load heat load cold load time series andlighting load series which constitutes totally 42 nodes so theinput layer dimension is set to 42 Since this paper simul-taneously predicts six attributes of wind power photovol-taic electricity heat and cold load the output layer has 5nodes that is the output layer dimension is set to 5 In thispaper based on experience and multiple trial and com-parison the number of hidden layers is set to 2 and thedimension of the hidden layer is set to 30 +e selection ofthe activation function is as shown in Figure 4(a) thebatch_size is set to 42 the numbers of iterations of an epochis set to 8784 the learning rate is set to 080 and the learningdelay rate is set to 005 +e network training function usesthe BPTT algorithm mentioned above

55 Evaluation Indicators Multienergy-load integrationforecasting simultaneously trains multiple forecastingtasks To comprehensively measure the forecasting effect itis necessary to evaluate the integrated forecasting accuracyas a whole [39] In this paper the average accuracyevaluation index based on weight coefficient is proposedFor the importance degree of different energy on the twosides of the RIES the corresponding weight coefficient isgiven to different energy output types For an RIES theuncertainties of the source-side wind light and otheruncertain resources are large and have a great influence onthe stable operation of the system and the energy dispatchmanagement +erefore the source-side wind forecastingis set to a higher weight in addition as the power gridplays a leading role in the integrated energy system byvirtue of its perfect architecture and central hub advan-tages correspondingly the power load forecast is set to ahigher weight

+is paper proposes three calculation indicators namelymean absolute percentage error (MAPE) mean accuracy

x1 x2 xm

x1 x2 xn

x1 x2 xk

Hiddenlayer

Multitask learningOutputlayer

Inputlayer

Single-task learning

Wind powerhistorical data

Electrical loadhistorical data

Multisourcehistorical data

Figure 5 Comparison of single-task and multitasking learningbased on the LSTM network

6 Complexity

(MA) and weightedmean accuracy (WMA) as shown in thefollowing equation

MAPE 1n

1113944

n

i1

P(i) minus 1113954P(i)

P(i)

11138681113868111386811138681113868111386811138681113868

11138681113868111386811138681113868111386811138681113868times 100

MA 1 minus MAPE

WMA λ1MA1 + λ2MA2 + middot middot middot + λkMAk

(3)

In the equation P(i) and 1113954P(i) respectively representthe actual energy value and the predicted energy value at thefirst moment n is the sample data amount MAk representsthe prediction accuracy of the k prediction task and λk

represents the weight coefficient of the k prediction task Inthis paper the weight ratios of the energy predictions onboth sides of the source and the load are set as shown inTable 1

6 Simulation Analysis

61 Data Description +e experimental data in this paperare derived from the measured data provided by RIES in acertain area of Tianjin +e RIES consists of a power systema thermal system and a natural gas system +e source sideincludes energy supply equipment such as wind turbinesand photovoltaic power generation equipment Couplinglinks include energy conversion equipment such as cold andthermal power supply equipment electric boilers and gasboilers Energy demand mainly includes power load heatload and cooling load +e load side and the energy storageside contain devices such as electrical energy storage andthermal energy storage to improve the flexibility of thesystem In addition in order to prove the validity of thesource-load vertical information interaction the historicaldata of the lighting load reflecting the typical electricity

usage behavior of the user are also sampled All historicalenergy data sampling interval is 1 h data from July 2015 toJune 2016 are used as training set data and data from July2016 are used as test set data with 1 h as time step based onthe historical data of the first 7 days the energy supply andenergy demand on both sides of the source and the load inthe next day are predicted +e total 57168 samples weredivided into two parts (92 as the train set and 8 as the testset) +e forecasting models were trained by the train datasetand validated on the test dataset +e statistical informationof the above dataset is shown in Table 2

62 Analysis of Examples +is section analyzes the effect ofRIES multienergy information interactive energy prediction+e selected sunny day and rainy day of the test set wereanalyzed separately +e energy prediction results of windpower output photovoltaic output electric load heat loadand cold load on both sides of the source and the load in thetwo scenarios are shown in Figures 7 and 8 respectively

+e prediction results of Figures 7 and 8 show that theproposed source-load parallel LSTM learning model canobtain ideal results in wind energy photovoltaic electricload thermal load and cold load energy prediction in sunnyday scenario and rainy day scenario which confirms theeffectiveness of the model proposed in this paper +eprediction effect of each subtask in the parallel predictionresults shows a large difference Compared with the windand light prediction on the source side the predicted curvesof the charge heat and cold load on the charge side arecloser to the actual sequence curve which is due to theintermittent and random nature of source-side wind powerand photovoltaics

Comparing Figures 7(c)ndash7(e) it can be found that theelectric load is more volatile than the thermal load and thecold load which is because the thermal inertia and cold

LSTM 1

LSTMn

Forecast 1input data

Forecast ninput data

Parameters sharing

Forecast 1output data

Forecast noutput data

Inputssharing

Use

r beh

avio

r

Mul

tiene

rgy

sour

ce-lo

ad in

tegr

ated

fore

casti

ng re

sults

Figure 6 Interactive energy prediction model of RIES information based on parallel LSTM

Complexity 7

inertia of the thermal system are stronger [39] and thevolatility is also even smaller +ermal inertia means that thedynamic change of the thermal system is a slow process froma time scale Compared with the power system the heatenergy transmission exhibits a delay effect and the upperand lower fluctuations of the heat load exhibit an inertialeffect and the cold inertia is similar Because of the existenceof thermal inertia and cold inertia the thermal load in thethermal system is less random in the short-term predicted

time scale showing a characteristic of smooth variation soits prediction uncertainty is also small and the thermal loadand the cold load are relative to the electricity +e pre-diction error of the load is also relatively small ComparingFigures 7 and 8 it can be found that the uncertainty of actualphotovoltaic power and actual wind power in rainy dayscenes is larger than that of sunny days+e forecast curve ofphotovoltaic and wind power in sunny scenes is closer to thetrue value the electric load and heat load in rainy scenes

Table 1 Weight settings of source-load double-sided energy prediction tasks

Prediction task Wind power PV power Electric load Heat load Cold loadWeight coefficient 025 025 02 015 015

Table 2 +e description of simulation dataset

Data type Max (kW) Median (kW) Min (kW) Mean (kW) Std (kW) Train set size Test set sizeWind power 317561 233774 36612 216102 78964 8784 744Photovoltaic power 235000 0000 0000 42507 70557 8784 744Electric load 696585 584923 209872 523803 139174 8784 744Heat load 148473 92719 60532 100076 27240 8784 744Cold load 436964 223213 39046 221949 135304 8784 744Lighting load 03333 50000 2778 44792 27734 8784 744Multisource fusion mdash mdash mdash mdash mdash 52704 4464

True valuePredictive value

50

100

150

200

250

300

Win

d po

wer

(kW

)

5 10 15 20 250t (h)

(a)

True valuePredictive value

5 10 15 20 250t (h)

0

50

100

150

200

250

PV p

ower

(kW

)

(b)

True valuePredictive value

5 10 15 20 250t (h)

200

300

400

500

600

700

Elec

tric

load

(kW

)

(c)

True valuePredictive value

5 10 15 20 250t (h)

40

60

80

100

120

140

Hea

d lo

ad (k

W)

(d)

True valuePredictive value

5 10 15 20 250t (h)

0

100

200

300

400

500

Col

d lo

ad (k

W)

(e)

Figure 7 +e results of source-load bilateral energy prediction within a sunny day (a) Prediction of wind power (b) Prediction of PVpower (c) Prediction of electric load (d) Prediction of heat load (e) Prediction of cold load

8 Complexity

should increase as a whole +is is due to the cold weather inthe rainy scenes which affects the userrsquos energy use behaviorand increases the userrsquos demand for electricity and heat thusincreasing the electrical load and heat load

621 Comparative Analysis of Multienergy Load InformationInteraction In order to verify the effect of interactivelearning of horizontal and vertical multienergy informationthe parallel LSTM predictions considering user energy be-havior and parallel LSTM prediction without consideringuser energy behavior and the results of single LSTM pre-diction without considering user energy behavior are pro-posed to conduct a comparative analysis In order to ensurethe fairness of the experimental test this paper uses the samedataset for training and testing

+e comparison experiment setup is shown in Table 3 Inthe parallel LSTM learning without considering the userrsquosenergy behavior the input sample is based on the originalprediction model A and the lighting load data reflecting theuser behavior in the input sample are not considered +eprediction is based only on the historical data of the pre-dicted object the model C does not consider the singleLSTMprediction of the userrsquos energy behavior and the inputand output of each subprediction model only consider theinput and output of a single energy and there is also noparameter sharing between the LSTM unit of the different

submodels Model B uses parallel LSTM learning +ere isparameter sharing among each subtask which belongs to amultitask learning Model C uses a single LSTM learningand each subtask does not affect each other It belongs to asingle-task learning

+e prediction accuracy of the multienergy charge in-formation interaction comparison experiment is shown inTable 4 and Figure 9 +rough the analysis of Table 4 andFigure 9 it shows that the average prediction accuracy of theweights of the models A B and C are 09009 08917 and08728 respectively +e prediction result of the model A isthe best the model B is the second and the model C is thethird +at is considering the interaction of vertical andhorizontal information the energy predictions on both sidesof the source and the load show better prediction accuracy

Comparing the prediction results of Model B and ModelC for the five subprediction tasks when the different qualityenergies are predicted together in parallel a higher precisionthan the single prediction is obtained It can be seen thatparallel LSTM learning considering horizontal multienergyinformation interaction utilizes the coupled data betweendifferent systems to fully exploit the interactive informationbetween different quality energy sources of the integratedenergy system so that the prediction performance of acertain energy can be improved In addition considering themultienergy horizontal information interaction the pre-dicted accuracy of the load-side electrical load heat load and

True valuePredictive value

5 10 15 20 250t (h)

100

150

200

250

300

350W

ind

pow

er (k

W)

(a)

True valuePredictive value

5 10 15 20 250t (h)

0

20

40

60

80

100

120

PV p

ower

(kW

)(b)

True valuePredictive value

5 10 15 20 250t (h)

300

400

500

600

700

Elec

tric

load

(kW

)

(c)

True valuePredictive value

5 10 15 20 250t (h)

60

80

100

120

140

160

Hea

t loa

d (k

W)

(d)

True valuePredictive value

5 10 15 20 250t (h)

0

100

200

300

400

500

Col

d lo

ad (k

W)

(e)

Figure 8+e results of source-load bilateral energy prediction within a rainy day (a) Prediction of wind power (b) Prediction of PV power(c) Prediction of electric load (d) Prediction of heat load (e) Prediction of cold load

Complexity 9

cold load are increased by 313 253 and 267 re-spectively while the predicted accuracy of the source-sidewind power and photovoltaic is increased by 152 and146 and the energy prediction accuracy of the charge sideis relatively high It can be seen that the source-side un-certainty resources are also affected by other energy in-teractions in the system but the impact is small while thecharge side demand is largely affected by other energy in-teractions in the system And the information interaction ismore intimate between each other

Comparing the prediction results of Model A and ModelB when the userrsquos energy behavior information interactsbetween the source and the load in vertical direction theaverage weight accuracy of the overall prediction is increasedby 103 Comparing the performance improvement effectsof the five subprediction tasks it can be found that thepercentage improvement of PV prediction accuracy is185 and the improvement effect is relatively best fol-lowed by the electrical load while the prediction accuracy ofwind power heat load and cooling load is not obvious Itcan be seen that considering the source-load vertical in-formation interaction can improve the energy predictionaccuracy on both sides of the source and the load but thespecific user energy behavior has different differences in theenergy prediction of different types of users and the user

lighting behavior is obviously related to photovoltaic powergeneration and user electricity +e interaction betweenloads is more tight and its effect on PV prediction andelectric load forecasting is more obvious

622 Comparison of Deep Neural Network Algorithms withOther Algorithms In order to verify the prediction effect ofthe LSTM deep neural network algorithm the LSTM modeland the autoregressive integrated moving average model(ARIMA) algorithm and the classic BP neural network(BPNN) algorithm in this paper are proposed +e simu-lation results of them were compared and analyzed In orderto ensure the fairness of the experimental test this paper usesthe same dataset for training and testing and both use asingle-task prediction strategy +e prediction results of thethree algorithms are shown in Table 5 and Figure 10

Observing the prediction results the prediction effect ofARIMA algorithm is the most general and the predictionaccuracy of BPNN algorithm is much higher than that ofARIMA algorithm +e prediction accuracy of each task ofLSTM deep learning prediction algorithm proposed in thispaper is the highest +e ARIMA algorithm can not effec-tively use the historical information in the long-term se-quence because it only uses a part of the historicalinformation to train which leads to the large predictionerror +e LSTM algorithm can fully utilize its strongnonlinear learning ability and long-term and short-termmemory advantages to exploit the information value of long-term historical data so the grasp of the energy prediction lawis more precise

In addition comparing the energy prediction results ofthe LSTM algorithm and the BPNN algorithm one by onethe LSTM algorithm predicts that the accuracy of thethermal load and the cold load prediction is more obvious Itcan be seen that due to the greater inertia of the thermalsystem the time scale of the change law is longer whichmakes the prediction accuracy of the long-term and short-term deep learning on the heat load and the cold loadsignificantly higher than the electric load

In order to evaluate the computational efficiency ofalgorithms four forecasting scenarios are performed tentimes so that the forecasting results can reliably be sum-marized as in Table 6 Parallel LSTM and Single LSTM havelonger training time but the prediction time is kept within

Table 3 Contrast experiment setting of multisource source information interaction

Model Vertical interaction Horizontal interactionModel A User energy behavior is considered Parallel LSTM learningModel B User energy behavior is not considered Parallel LSTM learningModel C User energy behavior is not considered Single-LSTM learning

Table 4 +ree kinds of model prediction results of information interaction comparison experiment

Precision Wind power PV power Electric load Heat load Cold load Mean precisionModel A 08602 08735 09236 09445 09407 09009Model B 08542 08576 09113 09401 09368 08917Model C 08414 08453 08836 09169 09124 08728

Model AModel BModel C

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

PV Electric Heat Cold MeanWindPrediction task

Figure 9 Comparison of prediction effects of three models

10 Complexity

one minute and can meet the requirements of short-termforecasting Further analysis shows that the parallel trainingtime and prediction time of parallel LSTM is greatly im-proved compared to single LSTM due to the combination offive subtasks into one multitask learning In summary al-though the LSTM algorithm is time consuming in trainingtime for short-term prediction the training process doesnot require online completion and the prediction time ofthe LSTM algorithm especially parallel LSTM can fullymeet the requirements of short-term prediction Consid-ering the prediction accuracy the sacrifice of the LSTMalgorithm in computing time is worthwhile

7 Conclusion

In this paper the horizontal and vertical information interactioncharacteristics between multiple energy loads in RIES areconsidered Using LSTM network deep learning algorithm a

parallel LSTM-based multienergy-load information interactiveshort-term energy prediction method is proposed+rough thesimulation of the test data the following conclusions are drawn

(1) +e ldquoinformation interactionrdquo strategy proposed inthis paper can effectively improve the predictionperformance of source-load energy prediction inregional integrated energy systems

(2) Parallel prediction considering horizontal in-teraction mainly affects the prediction effect on theload side while the information sharing of user il-lumination behavior considering vertical interactionmainly affects the improvement of PV predictioneffect

(3) Compared with the conventional prediction algo-rithm the LSTM deep learning algorithm can learnthe deep hidden information of historical dataimprove the prediction accuracy while meetingcalculation time and prove the effectiveness of thealgorithm in the prediction of time series featuresequence

However the performance of the proposed parallelLSTM algorithm may be further enhanced by some evo-lutionary algorithms or multiobjective optimization algo-rithms such as differential evolution algorithm [40]multiobjective extremal optimization [41] and evolutionarymultiobjective optimization algorithms [42ndash47] Of coursesome significant topics are worth researching on the opti-mization of the proposed parallel LSTM by multiobjectiveoptimization

Data Availability

+e data used to support the findings of this study areavailable from the corresponding author upon request

Conflicts of Interest

+e authors declare no potential conflicts of interest

Acknowledgments

+is work was fully supported by the National Key Researchand Development Program of China (no 2018YFB0904200)

References

[1] A Q Huang M L Crow G T Heydt J P Zheng andS J Dale ldquo+e future renewable electric energy delivery andmanagement (FREEDM) system the energy internetrdquo Pro-ceedings of the IEEE vol 99 no 1 pp 133ndash148 2011

[2] I Colak S Sagiroglu G Fulli M Yesilbudak and C CovrigldquoA survey on the critical issues in smart grid technologiesrdquo

Table 5 Comparison of prediction results of three algorithms

Precision Wind power PV power Electric load Heat load Cold load Mean precisionLSTM 08414 08453 08836 09169 09124 08728BPNN 08216 08314 08695 08912 08889 08542ARMIA 07733 07824 08282 08454 08408 08075

Prediction precision ofwind powerPrediction precision ofPV powerPrediction precision ofelectric load

Prediction precision ofhead loadPrediction precision ofcold loadWeight average predictionprecision

07

075

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

BPNN ARMIALSTMPrediction algorithm

Figure 10 Comparison of prediction effects of three algorithms

Table 6 Comparison of computational efficiency of fouralgorithms

Algorithm Training timeaverage (sec)

Prediction timeaverage (sec)

Parallel LSTM 436374 1002Single LSTM 1218337 3662BPNN 1094911 2253ARMIA 54696 215

Complexity 11

Renewable and Sustainable Energy Reviews vol 54 pp 396ndash405 2016

[3] C Tu X He Z Shuai and F Jiang ldquoBig data issues in smartgrid-a reviewrdquo Renewable and Sustainable Energy Reviewsvol 79 pp 1099ndash1107 2017

[4] Y Zhang R Yang K Zhang H Jiang and J J ZhangldquoConsumption behavior analytics-aided energy forecastingand dispatchrdquo IEEE Intelligent Systems vol 32 no 4pp 59ndash63 2017

[5] J Deng F Wang Y Chen and X Zhao ldquoFrom industry 40 toenergy 50 the concept connotation and system framework ofintelligent energy systemsrdquo Journal of Automation vol 41no 12 pp 2003ndash2016 2015

[6] Q Sun L Shi Y Ni D Si and J Zhu ldquoAn enhanced cas-cading failure model integrating data mining techniquerdquoProtection and Control of Modern Power Systems vol 2 2017

[7] D Zheng A T Eseye J Zhang and H Li ldquoShort-term windpower forecasting using a double-stage hierarchical ANFISapproach for energy management in microgridsrdquo Protectionand Control of Modern Power Systems vol 2 pp 1ndash10 2017

[8] Z Li L Ye Y Zhao X Song J Teng and J Jin ldquoShort-termwind power prediction based on extreme learning machinewith error correctionrdquo Protection and Control of ModernPower Systems vol 1 no 1 pp 1ndash8 2016

[9] H Z Wang Z X Lei X Zhang B Zhou and J C Peng ldquoAreview of deep learning for renewable energy forecastingrdquoEnergy Conversion and Management vol 198 pp 1ndash16 2019

[10] H Z Wang G B Wang G Q Li J C Peng and Y T LiuldquoDeep belief network based deterministic and probabilisticwind speed forecasting approachrdquo Applied Energy vol 182pp 80ndash93 2016

[11] M-R Chen G-Q Zeng K-D Lu and J Weng ldquoA two-layernonlinear combination method for short-term wind speedprediction based on ELM ENN and LSTMrdquo IEEE Internet ofings Journal vol 6 no 4 pp 6997ndash7010 2019

[12] H Liu X Mi and Y Li ldquoSmart multi-step deep learningmodel for wind speed forecasting based on variational modedecomposition singular spectrum analysis LSTM networkand ELMrdquo Energy Conversion and Management vol 159pp 54ndash64 2018

[13] J Chen G-Q Zeng W Zhou W Du and K-D Lu ldquoWindspeed forecasting using nonlinear-learning ensemble of deeplearning time series prediction and extremal optimizationrdquoEnergy Conversion and Management vol 165 pp 681ndash6952018

[14] W Kong Z Y Dong Y Jia D J Hill Y Xu and Y ZhangldquoShort-term residential load forecasting based on LSTM re-current neural networkrdquo IEEE Transactions on Smart Gridvol 10 no 1 2017

[15] H Gao W Huang X Yang Y Duan and Y Yin ldquoTowardservice selection for workflow reconfiguration an interface-based computing solutionrdquo Future Generation ComputerSystems vol 87 pp 298ndash311 2018

[16] L Qi Q He F Chen et al ldquoFinding all you need web APIsrecommendation in web of things through keywords searchrdquoIEEE Transactions on Computational Social Systems vol 6no 5 2019

[17] L Qi R Wang C Hu S Li Q He and X Xu ldquoTime-awaredistributed service recommendation with privacy-preserva-tionrdquo Information Sciences vol 480 pp 354ndash364 2019

[18] Y Zhang and Q Yang ldquoAn overview of multi-task learningrdquoNational Science Review vol 5 no 1 pp 34ndash47 2018

[19] A Argyriou M T Evgeniou and M Pontil ldquoConvex multi-task feature learningrdquo Machine Learning vol 73 no 3pp 243ndash272 2008

[20] T Jebara ldquoMultitask sparsity via maximum entropy dis-criminationrdquo Journal of Machine Learning Research vol 12no 1 pp 75ndash110 2011

[21] L Qi Y Chen Y Yuan S Fu X Zhang and X Xu ldquoA QoS-aware virtual machine scheduling method for energy con-servation in cloud-based cyber-physical systemsrdquo WorldWide Web vol 5 pp 1ndash23 2019

[22] R Zhang P Xie C Wang G Liu and S Wan ldquoClassifyingtransportation mode and speed from trajectory data via deepmulti-scale learningrdquo Computer Networks vol 162 p 1068612019

[23] W Wang D Wang H Jia et al ldquoReview of steady-stateanalysis of typical regional integrated energy system under thebackground of energy internetrdquo Proceedings of the CSEE12pp 3292ndash3306 2016

[24] S Ding S Qu Y Xi and S Wan ldquoStimulus-driven andconcept-driven analysis for image caption generationrdquoNeurocomputing 2019

[25] P Vrba V Marik P Siano et al ldquoA review of agent andservice-oriented concepts applied to intelligent energy sys-temsrdquo IEEE Transactions on Industrial Informatics vol 10no 3 pp 1890ndash1903 2014

[26] W Jia C Kang and Q Chen ldquoAnalysis on demand-sideinteractive response capability for power system dispatch in asmart grid frameworkrdquo Electric Power Systems Researchvol 90 pp 11ndash17 2012

[27] Z C Lipton J Berkowitz and C Elkan ldquoA critical review ofrecurrent neural networks for sequence learningrdquo CoRR2015 httpsarxivorgpdf150600019pdf

[28] T N Sainath O Vinyals A Senior and H Sak ldquoCon-volutional long short-term memory fully connected deepneural networksrdquo in Proceedings of the 2015 IEEE In-ternational Conference on Acoustics Speech and Signal Pro-cessing (ICASSP) pp 4580ndash4584 IEEE Brisbane AustraliaApril 2015

[29] Y Guo Z Cheng L Nie YWang J Ma andM KankanhallildquoAttentive long short-term preference modeling for person-alized product searchrdquo ACM Transactions on InformationSystems (TOIS) vol 37 no 2 p 19 2019

[30] A Graves ldquoLong short-term memoryrdquo Neural Computationvol 9 pp 1735ndash1780 1997

[31] F A Gers J Schmidhuber and F Cummins ldquoLearning toforget continual prediction with LSTMrdquo Neural Computa-tion vol 12 no 10 pp 2451ndash2471 2000

[32] M L Seltzer and J Droppo ldquoMulti-task learning in deepneural networks for improved phoneme recognitionrdquo inProceedings of the 2013 IEEE International Conference onAcoustics Speech and Signal Processing pp 6965ndash6969 IEEEVancouver Canada May 2013

[33] S Wan Y Zhao T Wang Z Gu Q H Abbasi andK-K R Choo ldquoMulti-dimensional data indexing and rangequery processing via Voronoi diagram for internet of thingsrdquoFuture Generation Computer Systems vol 91 pp 382ndash3912019

[34] Z Gao D Y Wang S H Wan H Zhang and Y L WangldquoCognitive-inspired class-statistic matching with triple-con-strain for camera free 3D object retrievalrdquo Future GenerationComputer Systems vol 94 pp 641ndash653 2019

[35] C A Kang A R Brandt and L J Durlofsky ldquoOptimaloperation of an integrated energy system including fossil fuel

12 Complexity

power generation CO2 capture and windrdquo Energy vol 36no 12 pp 6806ndash6820 2011

[36] Q Zhang SWan BWang DW Gao and HMa ldquoAnomalydetection based on randommatrix theory for industrial powersystemsrdquo Journal of Systems Architecture vol 95 pp 67ndash742019

[37] Z Gao H Z Xuan H Zhang S Wan and K K R ChooldquoAdaptive fusion and category-level dictionary learningmodel for multi-view human action recognitionrdquo IEEE In-ternet of ings Journal p 1 2019

[38] WWu K Chen Y Qiao and Z Lu ldquoProbabilistic short-termwind power forecasting based on deep neural networksrdquo inProceedings of the 2016 International Conference on Proba-bilistic Methods Applied to Power Systems (PMAPS) pp 1ndash8Beijing China October 2016

[39] K M Powell A Sriprasad W J Cole and T F EdgarldquoHeating cooling and electrical load forecasting for a large-scale district energy systemrdquo Energy vol 74 pp 877ndash8852014

[40] A Peimankar S J Weddell T Jalal and A C LapthornldquoMulti-objective ensemble forecasting with an application topower transformersrdquo Applied Soft Computing vol 68pp 233ndash248 2018

[41] G-Q Zeng J Chen L-M Li et al ldquoAn improved multi-objective population-based extremal optimization algorithmwith polynomial mutationrdquo Information Sciences vol 330pp 49ndash73 2016

[42] Y-L Hu and L Chen ldquoA nonlinear hybrid wind speedforecasting model using LSTM network hysteretic ELM andDifferential Evolution algorithmrdquo Energy Conversion andManagement vol 173 pp 123ndash142 2018

[43] G-Q Zeng J Chen Y-X Dai L-M Li C-W Zheng andM-R Chen ldquoAn improved Design of fractional order PIDcontroller for automatic regulator voltage system based onmulti-objective extremal optimizationrdquo Neurocomputingvol 160 pp 173ndash184 2015

[44] X Xu R Gu F Dai L Qi and S Wan ldquoMulti-objectivecomputation offloading for internet of vehicles in cloud-edgecomputingrdquo Wireless Networks pp 1ndash19 2019

[45] S Wan X Li Y Xue W Lin and X Xu ldquoEfficient com-putation offloading for internet of vehicles in edge comput-ing-assisted 5G networksrdquo e Journal of Supercomputingpp 1ndash30 2019

[46] B Wang S Wan X Zhang and K-K R Choo ldquoA novelindex for assessing the robustness of integrated electricalnetwork and a natural gas networkrdquo IEEE Access vol 6pp 40400ndash40410 2018

[47] Y Yin F Yu Y Xu L Yu and J Mu ldquoNetwork location-aware service recommendation with random walk in cyber-physical systemsrdquo Sensors vol 17 no 9 p 2059 2017

Complexity 13

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Page 6: Parallel LSTM-Based Regional Integrated Energy System ...downloads.hindawi.com/journals/complexity/2019/7414318.pdf · Characteristics of Regional Integrated Energy System “Information

predictionrdquo and utilizes the strong nonlinear learning abilityand historical memory advantages of LSTM deep learningnetwork to conduct integrated prediction of uncertain re-sources at both source and load As shown in Figure 6 thehorizontal information interaction based on multienergyconversion information sharingmainly utilizes themultitasklearning parameter sharing mechanism of the parallel LSTMprediction network the vertical information interactionbased on the user behavior information sharing mainly usesthe user behavior data to guide source-load multitask pre-diction as an additional unified input information Para-metric sharing is weight sharing which takes advantage ofthe correlation between subtasks Input sharing refers todata sharing Different subtasks share historical power datasuch as wind power photovoltaic power power loadthermal load and cooling load

52 Data Preprocessing +e training samples select the windpower generation time series on the source side the photo-voltaic power generation time series the charge side on the loadside the heat load the cold load time series and the illumi-nation load series data Considering the input and output rangeof the nonlinear activation function in the model in order toavoid neuron saturation the data are normalized and mappedbetween [minus 1 1] as shown in the following equation

xprime x minus xmax + xmin( 11138572

xmax minus xmin( 11138572 (2)

In the equation x is the actual input or output data andxmax and xmin are the maximum and minimum values of thevariable respectively

53 Task Construction +e essence of using the multitasklearning method for multienergy-load integration predictionlies in the close correlation of each prediction subtask [37]+e traditional single-task prediction method uses a singledimension of historical observations to predict its futureoutput and the model output is a single attribute +emultitask learning integration prediction uses the multidi-mensional data on both sides of the source and the load as a

common input to effectively utilize the sample On the otherhand through the shared hidden layer node divisionmethodthe five attributes of wind photovoltaic power generationelectricity heat and cold load are simultaneously predictedto play the role of inductive bias and improve predictionaccuracy

54 Model Training +e LSTM-based source-charge in-tegrated prediction model mainly needs to determine thefollowing parameters input layer dimension input layertime step hidden layer number each hidden layer di-mension and output layer dimension [38]

+e time step of the input layer is generally the length ofthe variable time series used for the source-load integrationprediction According to the periodic characteristics of theprediction object sequence the input layer time step is set to24 that is the historical data whose resolution ratio is 1 hduring the first seven days before the input is used forforecasting and it takes up one input node per day+e inputlayer dimension is the number of nodes +e input layernode of this paper contains six objects wind power gen-eration time series photovoltaic power generation timeseries charge side load heat load cold load time series andlighting load series which constitutes totally 42 nodes so theinput layer dimension is set to 42 Since this paper simul-taneously predicts six attributes of wind power photovol-taic electricity heat and cold load the output layer has 5nodes that is the output layer dimension is set to 5 In thispaper based on experience and multiple trial and com-parison the number of hidden layers is set to 2 and thedimension of the hidden layer is set to 30 +e selection ofthe activation function is as shown in Figure 4(a) thebatch_size is set to 42 the numbers of iterations of an epochis set to 8784 the learning rate is set to 080 and the learningdelay rate is set to 005 +e network training function usesthe BPTT algorithm mentioned above

55 Evaluation Indicators Multienergy-load integrationforecasting simultaneously trains multiple forecastingtasks To comprehensively measure the forecasting effect itis necessary to evaluate the integrated forecasting accuracyas a whole [39] In this paper the average accuracyevaluation index based on weight coefficient is proposedFor the importance degree of different energy on the twosides of the RIES the corresponding weight coefficient isgiven to different energy output types For an RIES theuncertainties of the source-side wind light and otheruncertain resources are large and have a great influence onthe stable operation of the system and the energy dispatchmanagement +erefore the source-side wind forecastingis set to a higher weight in addition as the power gridplays a leading role in the integrated energy system byvirtue of its perfect architecture and central hub advan-tages correspondingly the power load forecast is set to ahigher weight

+is paper proposes three calculation indicators namelymean absolute percentage error (MAPE) mean accuracy

x1 x2 xm

x1 x2 xn

x1 x2 xk

Hiddenlayer

Multitask learningOutputlayer

Inputlayer

Single-task learning

Wind powerhistorical data

Electrical loadhistorical data

Multisourcehistorical data

Figure 5 Comparison of single-task and multitasking learningbased on the LSTM network

6 Complexity

(MA) and weightedmean accuracy (WMA) as shown in thefollowing equation

MAPE 1n

1113944

n

i1

P(i) minus 1113954P(i)

P(i)

11138681113868111386811138681113868111386811138681113868

11138681113868111386811138681113868111386811138681113868times 100

MA 1 minus MAPE

WMA λ1MA1 + λ2MA2 + middot middot middot + λkMAk

(3)

In the equation P(i) and 1113954P(i) respectively representthe actual energy value and the predicted energy value at thefirst moment n is the sample data amount MAk representsthe prediction accuracy of the k prediction task and λk

represents the weight coefficient of the k prediction task Inthis paper the weight ratios of the energy predictions onboth sides of the source and the load are set as shown inTable 1

6 Simulation Analysis

61 Data Description +e experimental data in this paperare derived from the measured data provided by RIES in acertain area of Tianjin +e RIES consists of a power systema thermal system and a natural gas system +e source sideincludes energy supply equipment such as wind turbinesand photovoltaic power generation equipment Couplinglinks include energy conversion equipment such as cold andthermal power supply equipment electric boilers and gasboilers Energy demand mainly includes power load heatload and cooling load +e load side and the energy storageside contain devices such as electrical energy storage andthermal energy storage to improve the flexibility of thesystem In addition in order to prove the validity of thesource-load vertical information interaction the historicaldata of the lighting load reflecting the typical electricity

usage behavior of the user are also sampled All historicalenergy data sampling interval is 1 h data from July 2015 toJune 2016 are used as training set data and data from July2016 are used as test set data with 1 h as time step based onthe historical data of the first 7 days the energy supply andenergy demand on both sides of the source and the load inthe next day are predicted +e total 57168 samples weredivided into two parts (92 as the train set and 8 as the testset) +e forecasting models were trained by the train datasetand validated on the test dataset +e statistical informationof the above dataset is shown in Table 2

62 Analysis of Examples +is section analyzes the effect ofRIES multienergy information interactive energy prediction+e selected sunny day and rainy day of the test set wereanalyzed separately +e energy prediction results of windpower output photovoltaic output electric load heat loadand cold load on both sides of the source and the load in thetwo scenarios are shown in Figures 7 and 8 respectively

+e prediction results of Figures 7 and 8 show that theproposed source-load parallel LSTM learning model canobtain ideal results in wind energy photovoltaic electricload thermal load and cold load energy prediction in sunnyday scenario and rainy day scenario which confirms theeffectiveness of the model proposed in this paper +eprediction effect of each subtask in the parallel predictionresults shows a large difference Compared with the windand light prediction on the source side the predicted curvesof the charge heat and cold load on the charge side arecloser to the actual sequence curve which is due to theintermittent and random nature of source-side wind powerand photovoltaics

Comparing Figures 7(c)ndash7(e) it can be found that theelectric load is more volatile than the thermal load and thecold load which is because the thermal inertia and cold

LSTM 1

LSTMn

Forecast 1input data

Forecast ninput data

Parameters sharing

Forecast 1output data

Forecast noutput data

Inputssharing

Use

r beh

avio

r

Mul

tiene

rgy

sour

ce-lo

ad in

tegr

ated

fore

casti

ng re

sults

Figure 6 Interactive energy prediction model of RIES information based on parallel LSTM

Complexity 7

inertia of the thermal system are stronger [39] and thevolatility is also even smaller +ermal inertia means that thedynamic change of the thermal system is a slow process froma time scale Compared with the power system the heatenergy transmission exhibits a delay effect and the upperand lower fluctuations of the heat load exhibit an inertialeffect and the cold inertia is similar Because of the existenceof thermal inertia and cold inertia the thermal load in thethermal system is less random in the short-term predicted

time scale showing a characteristic of smooth variation soits prediction uncertainty is also small and the thermal loadand the cold load are relative to the electricity +e pre-diction error of the load is also relatively small ComparingFigures 7 and 8 it can be found that the uncertainty of actualphotovoltaic power and actual wind power in rainy dayscenes is larger than that of sunny days+e forecast curve ofphotovoltaic and wind power in sunny scenes is closer to thetrue value the electric load and heat load in rainy scenes

Table 1 Weight settings of source-load double-sided energy prediction tasks

Prediction task Wind power PV power Electric load Heat load Cold loadWeight coefficient 025 025 02 015 015

Table 2 +e description of simulation dataset

Data type Max (kW) Median (kW) Min (kW) Mean (kW) Std (kW) Train set size Test set sizeWind power 317561 233774 36612 216102 78964 8784 744Photovoltaic power 235000 0000 0000 42507 70557 8784 744Electric load 696585 584923 209872 523803 139174 8784 744Heat load 148473 92719 60532 100076 27240 8784 744Cold load 436964 223213 39046 221949 135304 8784 744Lighting load 03333 50000 2778 44792 27734 8784 744Multisource fusion mdash mdash mdash mdash mdash 52704 4464

True valuePredictive value

50

100

150

200

250

300

Win

d po

wer

(kW

)

5 10 15 20 250t (h)

(a)

True valuePredictive value

5 10 15 20 250t (h)

0

50

100

150

200

250

PV p

ower

(kW

)

(b)

True valuePredictive value

5 10 15 20 250t (h)

200

300

400

500

600

700

Elec

tric

load

(kW

)

(c)

True valuePredictive value

5 10 15 20 250t (h)

40

60

80

100

120

140

Hea

d lo

ad (k

W)

(d)

True valuePredictive value

5 10 15 20 250t (h)

0

100

200

300

400

500

Col

d lo

ad (k

W)

(e)

Figure 7 +e results of source-load bilateral energy prediction within a sunny day (a) Prediction of wind power (b) Prediction of PVpower (c) Prediction of electric load (d) Prediction of heat load (e) Prediction of cold load

8 Complexity

should increase as a whole +is is due to the cold weather inthe rainy scenes which affects the userrsquos energy use behaviorand increases the userrsquos demand for electricity and heat thusincreasing the electrical load and heat load

621 Comparative Analysis of Multienergy Load InformationInteraction In order to verify the effect of interactivelearning of horizontal and vertical multienergy informationthe parallel LSTM predictions considering user energy be-havior and parallel LSTM prediction without consideringuser energy behavior and the results of single LSTM pre-diction without considering user energy behavior are pro-posed to conduct a comparative analysis In order to ensurethe fairness of the experimental test this paper uses the samedataset for training and testing

+e comparison experiment setup is shown in Table 3 Inthe parallel LSTM learning without considering the userrsquosenergy behavior the input sample is based on the originalprediction model A and the lighting load data reflecting theuser behavior in the input sample are not considered +eprediction is based only on the historical data of the pre-dicted object the model C does not consider the singleLSTMprediction of the userrsquos energy behavior and the inputand output of each subprediction model only consider theinput and output of a single energy and there is also noparameter sharing between the LSTM unit of the different

submodels Model B uses parallel LSTM learning +ere isparameter sharing among each subtask which belongs to amultitask learning Model C uses a single LSTM learningand each subtask does not affect each other It belongs to asingle-task learning

+e prediction accuracy of the multienergy charge in-formation interaction comparison experiment is shown inTable 4 and Figure 9 +rough the analysis of Table 4 andFigure 9 it shows that the average prediction accuracy of theweights of the models A B and C are 09009 08917 and08728 respectively +e prediction result of the model A isthe best the model B is the second and the model C is thethird +at is considering the interaction of vertical andhorizontal information the energy predictions on both sidesof the source and the load show better prediction accuracy

Comparing the prediction results of Model B and ModelC for the five subprediction tasks when the different qualityenergies are predicted together in parallel a higher precisionthan the single prediction is obtained It can be seen thatparallel LSTM learning considering horizontal multienergyinformation interaction utilizes the coupled data betweendifferent systems to fully exploit the interactive informationbetween different quality energy sources of the integratedenergy system so that the prediction performance of acertain energy can be improved In addition considering themultienergy horizontal information interaction the pre-dicted accuracy of the load-side electrical load heat load and

True valuePredictive value

5 10 15 20 250t (h)

100

150

200

250

300

350W

ind

pow

er (k

W)

(a)

True valuePredictive value

5 10 15 20 250t (h)

0

20

40

60

80

100

120

PV p

ower

(kW

)(b)

True valuePredictive value

5 10 15 20 250t (h)

300

400

500

600

700

Elec

tric

load

(kW

)

(c)

True valuePredictive value

5 10 15 20 250t (h)

60

80

100

120

140

160

Hea

t loa

d (k

W)

(d)

True valuePredictive value

5 10 15 20 250t (h)

0

100

200

300

400

500

Col

d lo

ad (k

W)

(e)

Figure 8+e results of source-load bilateral energy prediction within a rainy day (a) Prediction of wind power (b) Prediction of PV power(c) Prediction of electric load (d) Prediction of heat load (e) Prediction of cold load

Complexity 9

cold load are increased by 313 253 and 267 re-spectively while the predicted accuracy of the source-sidewind power and photovoltaic is increased by 152 and146 and the energy prediction accuracy of the charge sideis relatively high It can be seen that the source-side un-certainty resources are also affected by other energy in-teractions in the system but the impact is small while thecharge side demand is largely affected by other energy in-teractions in the system And the information interaction ismore intimate between each other

Comparing the prediction results of Model A and ModelB when the userrsquos energy behavior information interactsbetween the source and the load in vertical direction theaverage weight accuracy of the overall prediction is increasedby 103 Comparing the performance improvement effectsof the five subprediction tasks it can be found that thepercentage improvement of PV prediction accuracy is185 and the improvement effect is relatively best fol-lowed by the electrical load while the prediction accuracy ofwind power heat load and cooling load is not obvious Itcan be seen that considering the source-load vertical in-formation interaction can improve the energy predictionaccuracy on both sides of the source and the load but thespecific user energy behavior has different differences in theenergy prediction of different types of users and the user

lighting behavior is obviously related to photovoltaic powergeneration and user electricity +e interaction betweenloads is more tight and its effect on PV prediction andelectric load forecasting is more obvious

622 Comparison of Deep Neural Network Algorithms withOther Algorithms In order to verify the prediction effect ofthe LSTM deep neural network algorithm the LSTM modeland the autoregressive integrated moving average model(ARIMA) algorithm and the classic BP neural network(BPNN) algorithm in this paper are proposed +e simu-lation results of them were compared and analyzed In orderto ensure the fairness of the experimental test this paper usesthe same dataset for training and testing and both use asingle-task prediction strategy +e prediction results of thethree algorithms are shown in Table 5 and Figure 10

Observing the prediction results the prediction effect ofARIMA algorithm is the most general and the predictionaccuracy of BPNN algorithm is much higher than that ofARIMA algorithm +e prediction accuracy of each task ofLSTM deep learning prediction algorithm proposed in thispaper is the highest +e ARIMA algorithm can not effec-tively use the historical information in the long-term se-quence because it only uses a part of the historicalinformation to train which leads to the large predictionerror +e LSTM algorithm can fully utilize its strongnonlinear learning ability and long-term and short-termmemory advantages to exploit the information value of long-term historical data so the grasp of the energy prediction lawis more precise

In addition comparing the energy prediction results ofthe LSTM algorithm and the BPNN algorithm one by onethe LSTM algorithm predicts that the accuracy of thethermal load and the cold load prediction is more obvious Itcan be seen that due to the greater inertia of the thermalsystem the time scale of the change law is longer whichmakes the prediction accuracy of the long-term and short-term deep learning on the heat load and the cold loadsignificantly higher than the electric load

In order to evaluate the computational efficiency ofalgorithms four forecasting scenarios are performed tentimes so that the forecasting results can reliably be sum-marized as in Table 6 Parallel LSTM and Single LSTM havelonger training time but the prediction time is kept within

Table 3 Contrast experiment setting of multisource source information interaction

Model Vertical interaction Horizontal interactionModel A User energy behavior is considered Parallel LSTM learningModel B User energy behavior is not considered Parallel LSTM learningModel C User energy behavior is not considered Single-LSTM learning

Table 4 +ree kinds of model prediction results of information interaction comparison experiment

Precision Wind power PV power Electric load Heat load Cold load Mean precisionModel A 08602 08735 09236 09445 09407 09009Model B 08542 08576 09113 09401 09368 08917Model C 08414 08453 08836 09169 09124 08728

Model AModel BModel C

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

PV Electric Heat Cold MeanWindPrediction task

Figure 9 Comparison of prediction effects of three models

10 Complexity

one minute and can meet the requirements of short-termforecasting Further analysis shows that the parallel trainingtime and prediction time of parallel LSTM is greatly im-proved compared to single LSTM due to the combination offive subtasks into one multitask learning In summary al-though the LSTM algorithm is time consuming in trainingtime for short-term prediction the training process doesnot require online completion and the prediction time ofthe LSTM algorithm especially parallel LSTM can fullymeet the requirements of short-term prediction Consid-ering the prediction accuracy the sacrifice of the LSTMalgorithm in computing time is worthwhile

7 Conclusion

In this paper the horizontal and vertical information interactioncharacteristics between multiple energy loads in RIES areconsidered Using LSTM network deep learning algorithm a

parallel LSTM-based multienergy-load information interactiveshort-term energy prediction method is proposed+rough thesimulation of the test data the following conclusions are drawn

(1) +e ldquoinformation interactionrdquo strategy proposed inthis paper can effectively improve the predictionperformance of source-load energy prediction inregional integrated energy systems

(2) Parallel prediction considering horizontal in-teraction mainly affects the prediction effect on theload side while the information sharing of user il-lumination behavior considering vertical interactionmainly affects the improvement of PV predictioneffect

(3) Compared with the conventional prediction algo-rithm the LSTM deep learning algorithm can learnthe deep hidden information of historical dataimprove the prediction accuracy while meetingcalculation time and prove the effectiveness of thealgorithm in the prediction of time series featuresequence

However the performance of the proposed parallelLSTM algorithm may be further enhanced by some evo-lutionary algorithms or multiobjective optimization algo-rithms such as differential evolution algorithm [40]multiobjective extremal optimization [41] and evolutionarymultiobjective optimization algorithms [42ndash47] Of coursesome significant topics are worth researching on the opti-mization of the proposed parallel LSTM by multiobjectiveoptimization

Data Availability

+e data used to support the findings of this study areavailable from the corresponding author upon request

Conflicts of Interest

+e authors declare no potential conflicts of interest

Acknowledgments

+is work was fully supported by the National Key Researchand Development Program of China (no 2018YFB0904200)

References

[1] A Q Huang M L Crow G T Heydt J P Zheng andS J Dale ldquo+e future renewable electric energy delivery andmanagement (FREEDM) system the energy internetrdquo Pro-ceedings of the IEEE vol 99 no 1 pp 133ndash148 2011

[2] I Colak S Sagiroglu G Fulli M Yesilbudak and C CovrigldquoA survey on the critical issues in smart grid technologiesrdquo

Table 5 Comparison of prediction results of three algorithms

Precision Wind power PV power Electric load Heat load Cold load Mean precisionLSTM 08414 08453 08836 09169 09124 08728BPNN 08216 08314 08695 08912 08889 08542ARMIA 07733 07824 08282 08454 08408 08075

Prediction precision ofwind powerPrediction precision ofPV powerPrediction precision ofelectric load

Prediction precision ofhead loadPrediction precision ofcold loadWeight average predictionprecision

07

075

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

BPNN ARMIALSTMPrediction algorithm

Figure 10 Comparison of prediction effects of three algorithms

Table 6 Comparison of computational efficiency of fouralgorithms

Algorithm Training timeaverage (sec)

Prediction timeaverage (sec)

Parallel LSTM 436374 1002Single LSTM 1218337 3662BPNN 1094911 2253ARMIA 54696 215

Complexity 11

Renewable and Sustainable Energy Reviews vol 54 pp 396ndash405 2016

[3] C Tu X He Z Shuai and F Jiang ldquoBig data issues in smartgrid-a reviewrdquo Renewable and Sustainable Energy Reviewsvol 79 pp 1099ndash1107 2017

[4] Y Zhang R Yang K Zhang H Jiang and J J ZhangldquoConsumption behavior analytics-aided energy forecastingand dispatchrdquo IEEE Intelligent Systems vol 32 no 4pp 59ndash63 2017

[5] J Deng F Wang Y Chen and X Zhao ldquoFrom industry 40 toenergy 50 the concept connotation and system framework ofintelligent energy systemsrdquo Journal of Automation vol 41no 12 pp 2003ndash2016 2015

[6] Q Sun L Shi Y Ni D Si and J Zhu ldquoAn enhanced cas-cading failure model integrating data mining techniquerdquoProtection and Control of Modern Power Systems vol 2 2017

[7] D Zheng A T Eseye J Zhang and H Li ldquoShort-term windpower forecasting using a double-stage hierarchical ANFISapproach for energy management in microgridsrdquo Protectionand Control of Modern Power Systems vol 2 pp 1ndash10 2017

[8] Z Li L Ye Y Zhao X Song J Teng and J Jin ldquoShort-termwind power prediction based on extreme learning machinewith error correctionrdquo Protection and Control of ModernPower Systems vol 1 no 1 pp 1ndash8 2016

[9] H Z Wang Z X Lei X Zhang B Zhou and J C Peng ldquoAreview of deep learning for renewable energy forecastingrdquoEnergy Conversion and Management vol 198 pp 1ndash16 2019

[10] H Z Wang G B Wang G Q Li J C Peng and Y T LiuldquoDeep belief network based deterministic and probabilisticwind speed forecasting approachrdquo Applied Energy vol 182pp 80ndash93 2016

[11] M-R Chen G-Q Zeng K-D Lu and J Weng ldquoA two-layernonlinear combination method for short-term wind speedprediction based on ELM ENN and LSTMrdquo IEEE Internet ofings Journal vol 6 no 4 pp 6997ndash7010 2019

[12] H Liu X Mi and Y Li ldquoSmart multi-step deep learningmodel for wind speed forecasting based on variational modedecomposition singular spectrum analysis LSTM networkand ELMrdquo Energy Conversion and Management vol 159pp 54ndash64 2018

[13] J Chen G-Q Zeng W Zhou W Du and K-D Lu ldquoWindspeed forecasting using nonlinear-learning ensemble of deeplearning time series prediction and extremal optimizationrdquoEnergy Conversion and Management vol 165 pp 681ndash6952018

[14] W Kong Z Y Dong Y Jia D J Hill Y Xu and Y ZhangldquoShort-term residential load forecasting based on LSTM re-current neural networkrdquo IEEE Transactions on Smart Gridvol 10 no 1 2017

[15] H Gao W Huang X Yang Y Duan and Y Yin ldquoTowardservice selection for workflow reconfiguration an interface-based computing solutionrdquo Future Generation ComputerSystems vol 87 pp 298ndash311 2018

[16] L Qi Q He F Chen et al ldquoFinding all you need web APIsrecommendation in web of things through keywords searchrdquoIEEE Transactions on Computational Social Systems vol 6no 5 2019

[17] L Qi R Wang C Hu S Li Q He and X Xu ldquoTime-awaredistributed service recommendation with privacy-preserva-tionrdquo Information Sciences vol 480 pp 354ndash364 2019

[18] Y Zhang and Q Yang ldquoAn overview of multi-task learningrdquoNational Science Review vol 5 no 1 pp 34ndash47 2018

[19] A Argyriou M T Evgeniou and M Pontil ldquoConvex multi-task feature learningrdquo Machine Learning vol 73 no 3pp 243ndash272 2008

[20] T Jebara ldquoMultitask sparsity via maximum entropy dis-criminationrdquo Journal of Machine Learning Research vol 12no 1 pp 75ndash110 2011

[21] L Qi Y Chen Y Yuan S Fu X Zhang and X Xu ldquoA QoS-aware virtual machine scheduling method for energy con-servation in cloud-based cyber-physical systemsrdquo WorldWide Web vol 5 pp 1ndash23 2019

[22] R Zhang P Xie C Wang G Liu and S Wan ldquoClassifyingtransportation mode and speed from trajectory data via deepmulti-scale learningrdquo Computer Networks vol 162 p 1068612019

[23] W Wang D Wang H Jia et al ldquoReview of steady-stateanalysis of typical regional integrated energy system under thebackground of energy internetrdquo Proceedings of the CSEE12pp 3292ndash3306 2016

[24] S Ding S Qu Y Xi and S Wan ldquoStimulus-driven andconcept-driven analysis for image caption generationrdquoNeurocomputing 2019

[25] P Vrba V Marik P Siano et al ldquoA review of agent andservice-oriented concepts applied to intelligent energy sys-temsrdquo IEEE Transactions on Industrial Informatics vol 10no 3 pp 1890ndash1903 2014

[26] W Jia C Kang and Q Chen ldquoAnalysis on demand-sideinteractive response capability for power system dispatch in asmart grid frameworkrdquo Electric Power Systems Researchvol 90 pp 11ndash17 2012

[27] Z C Lipton J Berkowitz and C Elkan ldquoA critical review ofrecurrent neural networks for sequence learningrdquo CoRR2015 httpsarxivorgpdf150600019pdf

[28] T N Sainath O Vinyals A Senior and H Sak ldquoCon-volutional long short-term memory fully connected deepneural networksrdquo in Proceedings of the 2015 IEEE In-ternational Conference on Acoustics Speech and Signal Pro-cessing (ICASSP) pp 4580ndash4584 IEEE Brisbane AustraliaApril 2015

[29] Y Guo Z Cheng L Nie YWang J Ma andM KankanhallildquoAttentive long short-term preference modeling for person-alized product searchrdquo ACM Transactions on InformationSystems (TOIS) vol 37 no 2 p 19 2019

[30] A Graves ldquoLong short-term memoryrdquo Neural Computationvol 9 pp 1735ndash1780 1997

[31] F A Gers J Schmidhuber and F Cummins ldquoLearning toforget continual prediction with LSTMrdquo Neural Computa-tion vol 12 no 10 pp 2451ndash2471 2000

[32] M L Seltzer and J Droppo ldquoMulti-task learning in deepneural networks for improved phoneme recognitionrdquo inProceedings of the 2013 IEEE International Conference onAcoustics Speech and Signal Processing pp 6965ndash6969 IEEEVancouver Canada May 2013

[33] S Wan Y Zhao T Wang Z Gu Q H Abbasi andK-K R Choo ldquoMulti-dimensional data indexing and rangequery processing via Voronoi diagram for internet of thingsrdquoFuture Generation Computer Systems vol 91 pp 382ndash3912019

[34] Z Gao D Y Wang S H Wan H Zhang and Y L WangldquoCognitive-inspired class-statistic matching with triple-con-strain for camera free 3D object retrievalrdquo Future GenerationComputer Systems vol 94 pp 641ndash653 2019

[35] C A Kang A R Brandt and L J Durlofsky ldquoOptimaloperation of an integrated energy system including fossil fuel

12 Complexity

power generation CO2 capture and windrdquo Energy vol 36no 12 pp 6806ndash6820 2011

[36] Q Zhang SWan BWang DW Gao and HMa ldquoAnomalydetection based on randommatrix theory for industrial powersystemsrdquo Journal of Systems Architecture vol 95 pp 67ndash742019

[37] Z Gao H Z Xuan H Zhang S Wan and K K R ChooldquoAdaptive fusion and category-level dictionary learningmodel for multi-view human action recognitionrdquo IEEE In-ternet of ings Journal p 1 2019

[38] WWu K Chen Y Qiao and Z Lu ldquoProbabilistic short-termwind power forecasting based on deep neural networksrdquo inProceedings of the 2016 International Conference on Proba-bilistic Methods Applied to Power Systems (PMAPS) pp 1ndash8Beijing China October 2016

[39] K M Powell A Sriprasad W J Cole and T F EdgarldquoHeating cooling and electrical load forecasting for a large-scale district energy systemrdquo Energy vol 74 pp 877ndash8852014

[40] A Peimankar S J Weddell T Jalal and A C LapthornldquoMulti-objective ensemble forecasting with an application topower transformersrdquo Applied Soft Computing vol 68pp 233ndash248 2018

[41] G-Q Zeng J Chen L-M Li et al ldquoAn improved multi-objective population-based extremal optimization algorithmwith polynomial mutationrdquo Information Sciences vol 330pp 49ndash73 2016

[42] Y-L Hu and L Chen ldquoA nonlinear hybrid wind speedforecasting model using LSTM network hysteretic ELM andDifferential Evolution algorithmrdquo Energy Conversion andManagement vol 173 pp 123ndash142 2018

[43] G-Q Zeng J Chen Y-X Dai L-M Li C-W Zheng andM-R Chen ldquoAn improved Design of fractional order PIDcontroller for automatic regulator voltage system based onmulti-objective extremal optimizationrdquo Neurocomputingvol 160 pp 173ndash184 2015

[44] X Xu R Gu F Dai L Qi and S Wan ldquoMulti-objectivecomputation offloading for internet of vehicles in cloud-edgecomputingrdquo Wireless Networks pp 1ndash19 2019

[45] S Wan X Li Y Xue W Lin and X Xu ldquoEfficient com-putation offloading for internet of vehicles in edge comput-ing-assisted 5G networksrdquo e Journal of Supercomputingpp 1ndash30 2019

[46] B Wang S Wan X Zhang and K-K R Choo ldquoA novelindex for assessing the robustness of integrated electricalnetwork and a natural gas networkrdquo IEEE Access vol 6pp 40400ndash40410 2018

[47] Y Yin F Yu Y Xu L Yu and J Mu ldquoNetwork location-aware service recommendation with random walk in cyber-physical systemsrdquo Sensors vol 17 no 9 p 2059 2017

Complexity 13

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Page 7: Parallel LSTM-Based Regional Integrated Energy System ...downloads.hindawi.com/journals/complexity/2019/7414318.pdf · Characteristics of Regional Integrated Energy System “Information

(MA) and weightedmean accuracy (WMA) as shown in thefollowing equation

MAPE 1n

1113944

n

i1

P(i) minus 1113954P(i)

P(i)

11138681113868111386811138681113868111386811138681113868

11138681113868111386811138681113868111386811138681113868times 100

MA 1 minus MAPE

WMA λ1MA1 + λ2MA2 + middot middot middot + λkMAk

(3)

In the equation P(i) and 1113954P(i) respectively representthe actual energy value and the predicted energy value at thefirst moment n is the sample data amount MAk representsthe prediction accuracy of the k prediction task and λk

represents the weight coefficient of the k prediction task Inthis paper the weight ratios of the energy predictions onboth sides of the source and the load are set as shown inTable 1

6 Simulation Analysis

61 Data Description +e experimental data in this paperare derived from the measured data provided by RIES in acertain area of Tianjin +e RIES consists of a power systema thermal system and a natural gas system +e source sideincludes energy supply equipment such as wind turbinesand photovoltaic power generation equipment Couplinglinks include energy conversion equipment such as cold andthermal power supply equipment electric boilers and gasboilers Energy demand mainly includes power load heatload and cooling load +e load side and the energy storageside contain devices such as electrical energy storage andthermal energy storage to improve the flexibility of thesystem In addition in order to prove the validity of thesource-load vertical information interaction the historicaldata of the lighting load reflecting the typical electricity

usage behavior of the user are also sampled All historicalenergy data sampling interval is 1 h data from July 2015 toJune 2016 are used as training set data and data from July2016 are used as test set data with 1 h as time step based onthe historical data of the first 7 days the energy supply andenergy demand on both sides of the source and the load inthe next day are predicted +e total 57168 samples weredivided into two parts (92 as the train set and 8 as the testset) +e forecasting models were trained by the train datasetand validated on the test dataset +e statistical informationof the above dataset is shown in Table 2

62 Analysis of Examples +is section analyzes the effect ofRIES multienergy information interactive energy prediction+e selected sunny day and rainy day of the test set wereanalyzed separately +e energy prediction results of windpower output photovoltaic output electric load heat loadand cold load on both sides of the source and the load in thetwo scenarios are shown in Figures 7 and 8 respectively

+e prediction results of Figures 7 and 8 show that theproposed source-load parallel LSTM learning model canobtain ideal results in wind energy photovoltaic electricload thermal load and cold load energy prediction in sunnyday scenario and rainy day scenario which confirms theeffectiveness of the model proposed in this paper +eprediction effect of each subtask in the parallel predictionresults shows a large difference Compared with the windand light prediction on the source side the predicted curvesof the charge heat and cold load on the charge side arecloser to the actual sequence curve which is due to theintermittent and random nature of source-side wind powerand photovoltaics

Comparing Figures 7(c)ndash7(e) it can be found that theelectric load is more volatile than the thermal load and thecold load which is because the thermal inertia and cold

LSTM 1

LSTMn

Forecast 1input data

Forecast ninput data

Parameters sharing

Forecast 1output data

Forecast noutput data

Inputssharing

Use

r beh

avio

r

Mul

tiene

rgy

sour

ce-lo

ad in

tegr

ated

fore

casti

ng re

sults

Figure 6 Interactive energy prediction model of RIES information based on parallel LSTM

Complexity 7

inertia of the thermal system are stronger [39] and thevolatility is also even smaller +ermal inertia means that thedynamic change of the thermal system is a slow process froma time scale Compared with the power system the heatenergy transmission exhibits a delay effect and the upperand lower fluctuations of the heat load exhibit an inertialeffect and the cold inertia is similar Because of the existenceof thermal inertia and cold inertia the thermal load in thethermal system is less random in the short-term predicted

time scale showing a characteristic of smooth variation soits prediction uncertainty is also small and the thermal loadand the cold load are relative to the electricity +e pre-diction error of the load is also relatively small ComparingFigures 7 and 8 it can be found that the uncertainty of actualphotovoltaic power and actual wind power in rainy dayscenes is larger than that of sunny days+e forecast curve ofphotovoltaic and wind power in sunny scenes is closer to thetrue value the electric load and heat load in rainy scenes

Table 1 Weight settings of source-load double-sided energy prediction tasks

Prediction task Wind power PV power Electric load Heat load Cold loadWeight coefficient 025 025 02 015 015

Table 2 +e description of simulation dataset

Data type Max (kW) Median (kW) Min (kW) Mean (kW) Std (kW) Train set size Test set sizeWind power 317561 233774 36612 216102 78964 8784 744Photovoltaic power 235000 0000 0000 42507 70557 8784 744Electric load 696585 584923 209872 523803 139174 8784 744Heat load 148473 92719 60532 100076 27240 8784 744Cold load 436964 223213 39046 221949 135304 8784 744Lighting load 03333 50000 2778 44792 27734 8784 744Multisource fusion mdash mdash mdash mdash mdash 52704 4464

True valuePredictive value

50

100

150

200

250

300

Win

d po

wer

(kW

)

5 10 15 20 250t (h)

(a)

True valuePredictive value

5 10 15 20 250t (h)

0

50

100

150

200

250

PV p

ower

(kW

)

(b)

True valuePredictive value

5 10 15 20 250t (h)

200

300

400

500

600

700

Elec

tric

load

(kW

)

(c)

True valuePredictive value

5 10 15 20 250t (h)

40

60

80

100

120

140

Hea

d lo

ad (k

W)

(d)

True valuePredictive value

5 10 15 20 250t (h)

0

100

200

300

400

500

Col

d lo

ad (k

W)

(e)

Figure 7 +e results of source-load bilateral energy prediction within a sunny day (a) Prediction of wind power (b) Prediction of PVpower (c) Prediction of electric load (d) Prediction of heat load (e) Prediction of cold load

8 Complexity

should increase as a whole +is is due to the cold weather inthe rainy scenes which affects the userrsquos energy use behaviorand increases the userrsquos demand for electricity and heat thusincreasing the electrical load and heat load

621 Comparative Analysis of Multienergy Load InformationInteraction In order to verify the effect of interactivelearning of horizontal and vertical multienergy informationthe parallel LSTM predictions considering user energy be-havior and parallel LSTM prediction without consideringuser energy behavior and the results of single LSTM pre-diction without considering user energy behavior are pro-posed to conduct a comparative analysis In order to ensurethe fairness of the experimental test this paper uses the samedataset for training and testing

+e comparison experiment setup is shown in Table 3 Inthe parallel LSTM learning without considering the userrsquosenergy behavior the input sample is based on the originalprediction model A and the lighting load data reflecting theuser behavior in the input sample are not considered +eprediction is based only on the historical data of the pre-dicted object the model C does not consider the singleLSTMprediction of the userrsquos energy behavior and the inputand output of each subprediction model only consider theinput and output of a single energy and there is also noparameter sharing between the LSTM unit of the different

submodels Model B uses parallel LSTM learning +ere isparameter sharing among each subtask which belongs to amultitask learning Model C uses a single LSTM learningand each subtask does not affect each other It belongs to asingle-task learning

+e prediction accuracy of the multienergy charge in-formation interaction comparison experiment is shown inTable 4 and Figure 9 +rough the analysis of Table 4 andFigure 9 it shows that the average prediction accuracy of theweights of the models A B and C are 09009 08917 and08728 respectively +e prediction result of the model A isthe best the model B is the second and the model C is thethird +at is considering the interaction of vertical andhorizontal information the energy predictions on both sidesof the source and the load show better prediction accuracy

Comparing the prediction results of Model B and ModelC for the five subprediction tasks when the different qualityenergies are predicted together in parallel a higher precisionthan the single prediction is obtained It can be seen thatparallel LSTM learning considering horizontal multienergyinformation interaction utilizes the coupled data betweendifferent systems to fully exploit the interactive informationbetween different quality energy sources of the integratedenergy system so that the prediction performance of acertain energy can be improved In addition considering themultienergy horizontal information interaction the pre-dicted accuracy of the load-side electrical load heat load and

True valuePredictive value

5 10 15 20 250t (h)

100

150

200

250

300

350W

ind

pow

er (k

W)

(a)

True valuePredictive value

5 10 15 20 250t (h)

0

20

40

60

80

100

120

PV p

ower

(kW

)(b)

True valuePredictive value

5 10 15 20 250t (h)

300

400

500

600

700

Elec

tric

load

(kW

)

(c)

True valuePredictive value

5 10 15 20 250t (h)

60

80

100

120

140

160

Hea

t loa

d (k

W)

(d)

True valuePredictive value

5 10 15 20 250t (h)

0

100

200

300

400

500

Col

d lo

ad (k

W)

(e)

Figure 8+e results of source-load bilateral energy prediction within a rainy day (a) Prediction of wind power (b) Prediction of PV power(c) Prediction of electric load (d) Prediction of heat load (e) Prediction of cold load

Complexity 9

cold load are increased by 313 253 and 267 re-spectively while the predicted accuracy of the source-sidewind power and photovoltaic is increased by 152 and146 and the energy prediction accuracy of the charge sideis relatively high It can be seen that the source-side un-certainty resources are also affected by other energy in-teractions in the system but the impact is small while thecharge side demand is largely affected by other energy in-teractions in the system And the information interaction ismore intimate between each other

Comparing the prediction results of Model A and ModelB when the userrsquos energy behavior information interactsbetween the source and the load in vertical direction theaverage weight accuracy of the overall prediction is increasedby 103 Comparing the performance improvement effectsof the five subprediction tasks it can be found that thepercentage improvement of PV prediction accuracy is185 and the improvement effect is relatively best fol-lowed by the electrical load while the prediction accuracy ofwind power heat load and cooling load is not obvious Itcan be seen that considering the source-load vertical in-formation interaction can improve the energy predictionaccuracy on both sides of the source and the load but thespecific user energy behavior has different differences in theenergy prediction of different types of users and the user

lighting behavior is obviously related to photovoltaic powergeneration and user electricity +e interaction betweenloads is more tight and its effect on PV prediction andelectric load forecasting is more obvious

622 Comparison of Deep Neural Network Algorithms withOther Algorithms In order to verify the prediction effect ofthe LSTM deep neural network algorithm the LSTM modeland the autoregressive integrated moving average model(ARIMA) algorithm and the classic BP neural network(BPNN) algorithm in this paper are proposed +e simu-lation results of them were compared and analyzed In orderto ensure the fairness of the experimental test this paper usesthe same dataset for training and testing and both use asingle-task prediction strategy +e prediction results of thethree algorithms are shown in Table 5 and Figure 10

Observing the prediction results the prediction effect ofARIMA algorithm is the most general and the predictionaccuracy of BPNN algorithm is much higher than that ofARIMA algorithm +e prediction accuracy of each task ofLSTM deep learning prediction algorithm proposed in thispaper is the highest +e ARIMA algorithm can not effec-tively use the historical information in the long-term se-quence because it only uses a part of the historicalinformation to train which leads to the large predictionerror +e LSTM algorithm can fully utilize its strongnonlinear learning ability and long-term and short-termmemory advantages to exploit the information value of long-term historical data so the grasp of the energy prediction lawis more precise

In addition comparing the energy prediction results ofthe LSTM algorithm and the BPNN algorithm one by onethe LSTM algorithm predicts that the accuracy of thethermal load and the cold load prediction is more obvious Itcan be seen that due to the greater inertia of the thermalsystem the time scale of the change law is longer whichmakes the prediction accuracy of the long-term and short-term deep learning on the heat load and the cold loadsignificantly higher than the electric load

In order to evaluate the computational efficiency ofalgorithms four forecasting scenarios are performed tentimes so that the forecasting results can reliably be sum-marized as in Table 6 Parallel LSTM and Single LSTM havelonger training time but the prediction time is kept within

Table 3 Contrast experiment setting of multisource source information interaction

Model Vertical interaction Horizontal interactionModel A User energy behavior is considered Parallel LSTM learningModel B User energy behavior is not considered Parallel LSTM learningModel C User energy behavior is not considered Single-LSTM learning

Table 4 +ree kinds of model prediction results of information interaction comparison experiment

Precision Wind power PV power Electric load Heat load Cold load Mean precisionModel A 08602 08735 09236 09445 09407 09009Model B 08542 08576 09113 09401 09368 08917Model C 08414 08453 08836 09169 09124 08728

Model AModel BModel C

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

PV Electric Heat Cold MeanWindPrediction task

Figure 9 Comparison of prediction effects of three models

10 Complexity

one minute and can meet the requirements of short-termforecasting Further analysis shows that the parallel trainingtime and prediction time of parallel LSTM is greatly im-proved compared to single LSTM due to the combination offive subtasks into one multitask learning In summary al-though the LSTM algorithm is time consuming in trainingtime for short-term prediction the training process doesnot require online completion and the prediction time ofthe LSTM algorithm especially parallel LSTM can fullymeet the requirements of short-term prediction Consid-ering the prediction accuracy the sacrifice of the LSTMalgorithm in computing time is worthwhile

7 Conclusion

In this paper the horizontal and vertical information interactioncharacteristics between multiple energy loads in RIES areconsidered Using LSTM network deep learning algorithm a

parallel LSTM-based multienergy-load information interactiveshort-term energy prediction method is proposed+rough thesimulation of the test data the following conclusions are drawn

(1) +e ldquoinformation interactionrdquo strategy proposed inthis paper can effectively improve the predictionperformance of source-load energy prediction inregional integrated energy systems

(2) Parallel prediction considering horizontal in-teraction mainly affects the prediction effect on theload side while the information sharing of user il-lumination behavior considering vertical interactionmainly affects the improvement of PV predictioneffect

(3) Compared with the conventional prediction algo-rithm the LSTM deep learning algorithm can learnthe deep hidden information of historical dataimprove the prediction accuracy while meetingcalculation time and prove the effectiveness of thealgorithm in the prediction of time series featuresequence

However the performance of the proposed parallelLSTM algorithm may be further enhanced by some evo-lutionary algorithms or multiobjective optimization algo-rithms such as differential evolution algorithm [40]multiobjective extremal optimization [41] and evolutionarymultiobjective optimization algorithms [42ndash47] Of coursesome significant topics are worth researching on the opti-mization of the proposed parallel LSTM by multiobjectiveoptimization

Data Availability

+e data used to support the findings of this study areavailable from the corresponding author upon request

Conflicts of Interest

+e authors declare no potential conflicts of interest

Acknowledgments

+is work was fully supported by the National Key Researchand Development Program of China (no 2018YFB0904200)

References

[1] A Q Huang M L Crow G T Heydt J P Zheng andS J Dale ldquo+e future renewable electric energy delivery andmanagement (FREEDM) system the energy internetrdquo Pro-ceedings of the IEEE vol 99 no 1 pp 133ndash148 2011

[2] I Colak S Sagiroglu G Fulli M Yesilbudak and C CovrigldquoA survey on the critical issues in smart grid technologiesrdquo

Table 5 Comparison of prediction results of three algorithms

Precision Wind power PV power Electric load Heat load Cold load Mean precisionLSTM 08414 08453 08836 09169 09124 08728BPNN 08216 08314 08695 08912 08889 08542ARMIA 07733 07824 08282 08454 08408 08075

Prediction precision ofwind powerPrediction precision ofPV powerPrediction precision ofelectric load

Prediction precision ofhead loadPrediction precision ofcold loadWeight average predictionprecision

07

075

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

BPNN ARMIALSTMPrediction algorithm

Figure 10 Comparison of prediction effects of three algorithms

Table 6 Comparison of computational efficiency of fouralgorithms

Algorithm Training timeaverage (sec)

Prediction timeaverage (sec)

Parallel LSTM 436374 1002Single LSTM 1218337 3662BPNN 1094911 2253ARMIA 54696 215

Complexity 11

Renewable and Sustainable Energy Reviews vol 54 pp 396ndash405 2016

[3] C Tu X He Z Shuai and F Jiang ldquoBig data issues in smartgrid-a reviewrdquo Renewable and Sustainable Energy Reviewsvol 79 pp 1099ndash1107 2017

[4] Y Zhang R Yang K Zhang H Jiang and J J ZhangldquoConsumption behavior analytics-aided energy forecastingand dispatchrdquo IEEE Intelligent Systems vol 32 no 4pp 59ndash63 2017

[5] J Deng F Wang Y Chen and X Zhao ldquoFrom industry 40 toenergy 50 the concept connotation and system framework ofintelligent energy systemsrdquo Journal of Automation vol 41no 12 pp 2003ndash2016 2015

[6] Q Sun L Shi Y Ni D Si and J Zhu ldquoAn enhanced cas-cading failure model integrating data mining techniquerdquoProtection and Control of Modern Power Systems vol 2 2017

[7] D Zheng A T Eseye J Zhang and H Li ldquoShort-term windpower forecasting using a double-stage hierarchical ANFISapproach for energy management in microgridsrdquo Protectionand Control of Modern Power Systems vol 2 pp 1ndash10 2017

[8] Z Li L Ye Y Zhao X Song J Teng and J Jin ldquoShort-termwind power prediction based on extreme learning machinewith error correctionrdquo Protection and Control of ModernPower Systems vol 1 no 1 pp 1ndash8 2016

[9] H Z Wang Z X Lei X Zhang B Zhou and J C Peng ldquoAreview of deep learning for renewable energy forecastingrdquoEnergy Conversion and Management vol 198 pp 1ndash16 2019

[10] H Z Wang G B Wang G Q Li J C Peng and Y T LiuldquoDeep belief network based deterministic and probabilisticwind speed forecasting approachrdquo Applied Energy vol 182pp 80ndash93 2016

[11] M-R Chen G-Q Zeng K-D Lu and J Weng ldquoA two-layernonlinear combination method for short-term wind speedprediction based on ELM ENN and LSTMrdquo IEEE Internet ofings Journal vol 6 no 4 pp 6997ndash7010 2019

[12] H Liu X Mi and Y Li ldquoSmart multi-step deep learningmodel for wind speed forecasting based on variational modedecomposition singular spectrum analysis LSTM networkand ELMrdquo Energy Conversion and Management vol 159pp 54ndash64 2018

[13] J Chen G-Q Zeng W Zhou W Du and K-D Lu ldquoWindspeed forecasting using nonlinear-learning ensemble of deeplearning time series prediction and extremal optimizationrdquoEnergy Conversion and Management vol 165 pp 681ndash6952018

[14] W Kong Z Y Dong Y Jia D J Hill Y Xu and Y ZhangldquoShort-term residential load forecasting based on LSTM re-current neural networkrdquo IEEE Transactions on Smart Gridvol 10 no 1 2017

[15] H Gao W Huang X Yang Y Duan and Y Yin ldquoTowardservice selection for workflow reconfiguration an interface-based computing solutionrdquo Future Generation ComputerSystems vol 87 pp 298ndash311 2018

[16] L Qi Q He F Chen et al ldquoFinding all you need web APIsrecommendation in web of things through keywords searchrdquoIEEE Transactions on Computational Social Systems vol 6no 5 2019

[17] L Qi R Wang C Hu S Li Q He and X Xu ldquoTime-awaredistributed service recommendation with privacy-preserva-tionrdquo Information Sciences vol 480 pp 354ndash364 2019

[18] Y Zhang and Q Yang ldquoAn overview of multi-task learningrdquoNational Science Review vol 5 no 1 pp 34ndash47 2018

[19] A Argyriou M T Evgeniou and M Pontil ldquoConvex multi-task feature learningrdquo Machine Learning vol 73 no 3pp 243ndash272 2008

[20] T Jebara ldquoMultitask sparsity via maximum entropy dis-criminationrdquo Journal of Machine Learning Research vol 12no 1 pp 75ndash110 2011

[21] L Qi Y Chen Y Yuan S Fu X Zhang and X Xu ldquoA QoS-aware virtual machine scheduling method for energy con-servation in cloud-based cyber-physical systemsrdquo WorldWide Web vol 5 pp 1ndash23 2019

[22] R Zhang P Xie C Wang G Liu and S Wan ldquoClassifyingtransportation mode and speed from trajectory data via deepmulti-scale learningrdquo Computer Networks vol 162 p 1068612019

[23] W Wang D Wang H Jia et al ldquoReview of steady-stateanalysis of typical regional integrated energy system under thebackground of energy internetrdquo Proceedings of the CSEE12pp 3292ndash3306 2016

[24] S Ding S Qu Y Xi and S Wan ldquoStimulus-driven andconcept-driven analysis for image caption generationrdquoNeurocomputing 2019

[25] P Vrba V Marik P Siano et al ldquoA review of agent andservice-oriented concepts applied to intelligent energy sys-temsrdquo IEEE Transactions on Industrial Informatics vol 10no 3 pp 1890ndash1903 2014

[26] W Jia C Kang and Q Chen ldquoAnalysis on demand-sideinteractive response capability for power system dispatch in asmart grid frameworkrdquo Electric Power Systems Researchvol 90 pp 11ndash17 2012

[27] Z C Lipton J Berkowitz and C Elkan ldquoA critical review ofrecurrent neural networks for sequence learningrdquo CoRR2015 httpsarxivorgpdf150600019pdf

[28] T N Sainath O Vinyals A Senior and H Sak ldquoCon-volutional long short-term memory fully connected deepneural networksrdquo in Proceedings of the 2015 IEEE In-ternational Conference on Acoustics Speech and Signal Pro-cessing (ICASSP) pp 4580ndash4584 IEEE Brisbane AustraliaApril 2015

[29] Y Guo Z Cheng L Nie YWang J Ma andM KankanhallildquoAttentive long short-term preference modeling for person-alized product searchrdquo ACM Transactions on InformationSystems (TOIS) vol 37 no 2 p 19 2019

[30] A Graves ldquoLong short-term memoryrdquo Neural Computationvol 9 pp 1735ndash1780 1997

[31] F A Gers J Schmidhuber and F Cummins ldquoLearning toforget continual prediction with LSTMrdquo Neural Computa-tion vol 12 no 10 pp 2451ndash2471 2000

[32] M L Seltzer and J Droppo ldquoMulti-task learning in deepneural networks for improved phoneme recognitionrdquo inProceedings of the 2013 IEEE International Conference onAcoustics Speech and Signal Processing pp 6965ndash6969 IEEEVancouver Canada May 2013

[33] S Wan Y Zhao T Wang Z Gu Q H Abbasi andK-K R Choo ldquoMulti-dimensional data indexing and rangequery processing via Voronoi diagram for internet of thingsrdquoFuture Generation Computer Systems vol 91 pp 382ndash3912019

[34] Z Gao D Y Wang S H Wan H Zhang and Y L WangldquoCognitive-inspired class-statistic matching with triple-con-strain for camera free 3D object retrievalrdquo Future GenerationComputer Systems vol 94 pp 641ndash653 2019

[35] C A Kang A R Brandt and L J Durlofsky ldquoOptimaloperation of an integrated energy system including fossil fuel

12 Complexity

power generation CO2 capture and windrdquo Energy vol 36no 12 pp 6806ndash6820 2011

[36] Q Zhang SWan BWang DW Gao and HMa ldquoAnomalydetection based on randommatrix theory for industrial powersystemsrdquo Journal of Systems Architecture vol 95 pp 67ndash742019

[37] Z Gao H Z Xuan H Zhang S Wan and K K R ChooldquoAdaptive fusion and category-level dictionary learningmodel for multi-view human action recognitionrdquo IEEE In-ternet of ings Journal p 1 2019

[38] WWu K Chen Y Qiao and Z Lu ldquoProbabilistic short-termwind power forecasting based on deep neural networksrdquo inProceedings of the 2016 International Conference on Proba-bilistic Methods Applied to Power Systems (PMAPS) pp 1ndash8Beijing China October 2016

[39] K M Powell A Sriprasad W J Cole and T F EdgarldquoHeating cooling and electrical load forecasting for a large-scale district energy systemrdquo Energy vol 74 pp 877ndash8852014

[40] A Peimankar S J Weddell T Jalal and A C LapthornldquoMulti-objective ensemble forecasting with an application topower transformersrdquo Applied Soft Computing vol 68pp 233ndash248 2018

[41] G-Q Zeng J Chen L-M Li et al ldquoAn improved multi-objective population-based extremal optimization algorithmwith polynomial mutationrdquo Information Sciences vol 330pp 49ndash73 2016

[42] Y-L Hu and L Chen ldquoA nonlinear hybrid wind speedforecasting model using LSTM network hysteretic ELM andDifferential Evolution algorithmrdquo Energy Conversion andManagement vol 173 pp 123ndash142 2018

[43] G-Q Zeng J Chen Y-X Dai L-M Li C-W Zheng andM-R Chen ldquoAn improved Design of fractional order PIDcontroller for automatic regulator voltage system based onmulti-objective extremal optimizationrdquo Neurocomputingvol 160 pp 173ndash184 2015

[44] X Xu R Gu F Dai L Qi and S Wan ldquoMulti-objectivecomputation offloading for internet of vehicles in cloud-edgecomputingrdquo Wireless Networks pp 1ndash19 2019

[45] S Wan X Li Y Xue W Lin and X Xu ldquoEfficient com-putation offloading for internet of vehicles in edge comput-ing-assisted 5G networksrdquo e Journal of Supercomputingpp 1ndash30 2019

[46] B Wang S Wan X Zhang and K-K R Choo ldquoA novelindex for assessing the robustness of integrated electricalnetwork and a natural gas networkrdquo IEEE Access vol 6pp 40400ndash40410 2018

[47] Y Yin F Yu Y Xu L Yu and J Mu ldquoNetwork location-aware service recommendation with random walk in cyber-physical systemsrdquo Sensors vol 17 no 9 p 2059 2017

Complexity 13

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Page 8: Parallel LSTM-Based Regional Integrated Energy System ...downloads.hindawi.com/journals/complexity/2019/7414318.pdf · Characteristics of Regional Integrated Energy System “Information

inertia of the thermal system are stronger [39] and thevolatility is also even smaller +ermal inertia means that thedynamic change of the thermal system is a slow process froma time scale Compared with the power system the heatenergy transmission exhibits a delay effect and the upperand lower fluctuations of the heat load exhibit an inertialeffect and the cold inertia is similar Because of the existenceof thermal inertia and cold inertia the thermal load in thethermal system is less random in the short-term predicted

time scale showing a characteristic of smooth variation soits prediction uncertainty is also small and the thermal loadand the cold load are relative to the electricity +e pre-diction error of the load is also relatively small ComparingFigures 7 and 8 it can be found that the uncertainty of actualphotovoltaic power and actual wind power in rainy dayscenes is larger than that of sunny days+e forecast curve ofphotovoltaic and wind power in sunny scenes is closer to thetrue value the electric load and heat load in rainy scenes

Table 1 Weight settings of source-load double-sided energy prediction tasks

Prediction task Wind power PV power Electric load Heat load Cold loadWeight coefficient 025 025 02 015 015

Table 2 +e description of simulation dataset

Data type Max (kW) Median (kW) Min (kW) Mean (kW) Std (kW) Train set size Test set sizeWind power 317561 233774 36612 216102 78964 8784 744Photovoltaic power 235000 0000 0000 42507 70557 8784 744Electric load 696585 584923 209872 523803 139174 8784 744Heat load 148473 92719 60532 100076 27240 8784 744Cold load 436964 223213 39046 221949 135304 8784 744Lighting load 03333 50000 2778 44792 27734 8784 744Multisource fusion mdash mdash mdash mdash mdash 52704 4464

True valuePredictive value

50

100

150

200

250

300

Win

d po

wer

(kW

)

5 10 15 20 250t (h)

(a)

True valuePredictive value

5 10 15 20 250t (h)

0

50

100

150

200

250

PV p

ower

(kW

)

(b)

True valuePredictive value

5 10 15 20 250t (h)

200

300

400

500

600

700

Elec

tric

load

(kW

)

(c)

True valuePredictive value

5 10 15 20 250t (h)

40

60

80

100

120

140

Hea

d lo

ad (k

W)

(d)

True valuePredictive value

5 10 15 20 250t (h)

0

100

200

300

400

500

Col

d lo

ad (k

W)

(e)

Figure 7 +e results of source-load bilateral energy prediction within a sunny day (a) Prediction of wind power (b) Prediction of PVpower (c) Prediction of electric load (d) Prediction of heat load (e) Prediction of cold load

8 Complexity

should increase as a whole +is is due to the cold weather inthe rainy scenes which affects the userrsquos energy use behaviorand increases the userrsquos demand for electricity and heat thusincreasing the electrical load and heat load

621 Comparative Analysis of Multienergy Load InformationInteraction In order to verify the effect of interactivelearning of horizontal and vertical multienergy informationthe parallel LSTM predictions considering user energy be-havior and parallel LSTM prediction without consideringuser energy behavior and the results of single LSTM pre-diction without considering user energy behavior are pro-posed to conduct a comparative analysis In order to ensurethe fairness of the experimental test this paper uses the samedataset for training and testing

+e comparison experiment setup is shown in Table 3 Inthe parallel LSTM learning without considering the userrsquosenergy behavior the input sample is based on the originalprediction model A and the lighting load data reflecting theuser behavior in the input sample are not considered +eprediction is based only on the historical data of the pre-dicted object the model C does not consider the singleLSTMprediction of the userrsquos energy behavior and the inputand output of each subprediction model only consider theinput and output of a single energy and there is also noparameter sharing between the LSTM unit of the different

submodels Model B uses parallel LSTM learning +ere isparameter sharing among each subtask which belongs to amultitask learning Model C uses a single LSTM learningand each subtask does not affect each other It belongs to asingle-task learning

+e prediction accuracy of the multienergy charge in-formation interaction comparison experiment is shown inTable 4 and Figure 9 +rough the analysis of Table 4 andFigure 9 it shows that the average prediction accuracy of theweights of the models A B and C are 09009 08917 and08728 respectively +e prediction result of the model A isthe best the model B is the second and the model C is thethird +at is considering the interaction of vertical andhorizontal information the energy predictions on both sidesof the source and the load show better prediction accuracy

Comparing the prediction results of Model B and ModelC for the five subprediction tasks when the different qualityenergies are predicted together in parallel a higher precisionthan the single prediction is obtained It can be seen thatparallel LSTM learning considering horizontal multienergyinformation interaction utilizes the coupled data betweendifferent systems to fully exploit the interactive informationbetween different quality energy sources of the integratedenergy system so that the prediction performance of acertain energy can be improved In addition considering themultienergy horizontal information interaction the pre-dicted accuracy of the load-side electrical load heat load and

True valuePredictive value

5 10 15 20 250t (h)

100

150

200

250

300

350W

ind

pow

er (k

W)

(a)

True valuePredictive value

5 10 15 20 250t (h)

0

20

40

60

80

100

120

PV p

ower

(kW

)(b)

True valuePredictive value

5 10 15 20 250t (h)

300

400

500

600

700

Elec

tric

load

(kW

)

(c)

True valuePredictive value

5 10 15 20 250t (h)

60

80

100

120

140

160

Hea

t loa

d (k

W)

(d)

True valuePredictive value

5 10 15 20 250t (h)

0

100

200

300

400

500

Col

d lo

ad (k

W)

(e)

Figure 8+e results of source-load bilateral energy prediction within a rainy day (a) Prediction of wind power (b) Prediction of PV power(c) Prediction of electric load (d) Prediction of heat load (e) Prediction of cold load

Complexity 9

cold load are increased by 313 253 and 267 re-spectively while the predicted accuracy of the source-sidewind power and photovoltaic is increased by 152 and146 and the energy prediction accuracy of the charge sideis relatively high It can be seen that the source-side un-certainty resources are also affected by other energy in-teractions in the system but the impact is small while thecharge side demand is largely affected by other energy in-teractions in the system And the information interaction ismore intimate between each other

Comparing the prediction results of Model A and ModelB when the userrsquos energy behavior information interactsbetween the source and the load in vertical direction theaverage weight accuracy of the overall prediction is increasedby 103 Comparing the performance improvement effectsof the five subprediction tasks it can be found that thepercentage improvement of PV prediction accuracy is185 and the improvement effect is relatively best fol-lowed by the electrical load while the prediction accuracy ofwind power heat load and cooling load is not obvious Itcan be seen that considering the source-load vertical in-formation interaction can improve the energy predictionaccuracy on both sides of the source and the load but thespecific user energy behavior has different differences in theenergy prediction of different types of users and the user

lighting behavior is obviously related to photovoltaic powergeneration and user electricity +e interaction betweenloads is more tight and its effect on PV prediction andelectric load forecasting is more obvious

622 Comparison of Deep Neural Network Algorithms withOther Algorithms In order to verify the prediction effect ofthe LSTM deep neural network algorithm the LSTM modeland the autoregressive integrated moving average model(ARIMA) algorithm and the classic BP neural network(BPNN) algorithm in this paper are proposed +e simu-lation results of them were compared and analyzed In orderto ensure the fairness of the experimental test this paper usesthe same dataset for training and testing and both use asingle-task prediction strategy +e prediction results of thethree algorithms are shown in Table 5 and Figure 10

Observing the prediction results the prediction effect ofARIMA algorithm is the most general and the predictionaccuracy of BPNN algorithm is much higher than that ofARIMA algorithm +e prediction accuracy of each task ofLSTM deep learning prediction algorithm proposed in thispaper is the highest +e ARIMA algorithm can not effec-tively use the historical information in the long-term se-quence because it only uses a part of the historicalinformation to train which leads to the large predictionerror +e LSTM algorithm can fully utilize its strongnonlinear learning ability and long-term and short-termmemory advantages to exploit the information value of long-term historical data so the grasp of the energy prediction lawis more precise

In addition comparing the energy prediction results ofthe LSTM algorithm and the BPNN algorithm one by onethe LSTM algorithm predicts that the accuracy of thethermal load and the cold load prediction is more obvious Itcan be seen that due to the greater inertia of the thermalsystem the time scale of the change law is longer whichmakes the prediction accuracy of the long-term and short-term deep learning on the heat load and the cold loadsignificantly higher than the electric load

In order to evaluate the computational efficiency ofalgorithms four forecasting scenarios are performed tentimes so that the forecasting results can reliably be sum-marized as in Table 6 Parallel LSTM and Single LSTM havelonger training time but the prediction time is kept within

Table 3 Contrast experiment setting of multisource source information interaction

Model Vertical interaction Horizontal interactionModel A User energy behavior is considered Parallel LSTM learningModel B User energy behavior is not considered Parallel LSTM learningModel C User energy behavior is not considered Single-LSTM learning

Table 4 +ree kinds of model prediction results of information interaction comparison experiment

Precision Wind power PV power Electric load Heat load Cold load Mean precisionModel A 08602 08735 09236 09445 09407 09009Model B 08542 08576 09113 09401 09368 08917Model C 08414 08453 08836 09169 09124 08728

Model AModel BModel C

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

PV Electric Heat Cold MeanWindPrediction task

Figure 9 Comparison of prediction effects of three models

10 Complexity

one minute and can meet the requirements of short-termforecasting Further analysis shows that the parallel trainingtime and prediction time of parallel LSTM is greatly im-proved compared to single LSTM due to the combination offive subtasks into one multitask learning In summary al-though the LSTM algorithm is time consuming in trainingtime for short-term prediction the training process doesnot require online completion and the prediction time ofthe LSTM algorithm especially parallel LSTM can fullymeet the requirements of short-term prediction Consid-ering the prediction accuracy the sacrifice of the LSTMalgorithm in computing time is worthwhile

7 Conclusion

In this paper the horizontal and vertical information interactioncharacteristics between multiple energy loads in RIES areconsidered Using LSTM network deep learning algorithm a

parallel LSTM-based multienergy-load information interactiveshort-term energy prediction method is proposed+rough thesimulation of the test data the following conclusions are drawn

(1) +e ldquoinformation interactionrdquo strategy proposed inthis paper can effectively improve the predictionperformance of source-load energy prediction inregional integrated energy systems

(2) Parallel prediction considering horizontal in-teraction mainly affects the prediction effect on theload side while the information sharing of user il-lumination behavior considering vertical interactionmainly affects the improvement of PV predictioneffect

(3) Compared with the conventional prediction algo-rithm the LSTM deep learning algorithm can learnthe deep hidden information of historical dataimprove the prediction accuracy while meetingcalculation time and prove the effectiveness of thealgorithm in the prediction of time series featuresequence

However the performance of the proposed parallelLSTM algorithm may be further enhanced by some evo-lutionary algorithms or multiobjective optimization algo-rithms such as differential evolution algorithm [40]multiobjective extremal optimization [41] and evolutionarymultiobjective optimization algorithms [42ndash47] Of coursesome significant topics are worth researching on the opti-mization of the proposed parallel LSTM by multiobjectiveoptimization

Data Availability

+e data used to support the findings of this study areavailable from the corresponding author upon request

Conflicts of Interest

+e authors declare no potential conflicts of interest

Acknowledgments

+is work was fully supported by the National Key Researchand Development Program of China (no 2018YFB0904200)

References

[1] A Q Huang M L Crow G T Heydt J P Zheng andS J Dale ldquo+e future renewable electric energy delivery andmanagement (FREEDM) system the energy internetrdquo Pro-ceedings of the IEEE vol 99 no 1 pp 133ndash148 2011

[2] I Colak S Sagiroglu G Fulli M Yesilbudak and C CovrigldquoA survey on the critical issues in smart grid technologiesrdquo

Table 5 Comparison of prediction results of three algorithms

Precision Wind power PV power Electric load Heat load Cold load Mean precisionLSTM 08414 08453 08836 09169 09124 08728BPNN 08216 08314 08695 08912 08889 08542ARMIA 07733 07824 08282 08454 08408 08075

Prediction precision ofwind powerPrediction precision ofPV powerPrediction precision ofelectric load

Prediction precision ofhead loadPrediction precision ofcold loadWeight average predictionprecision

07

075

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

BPNN ARMIALSTMPrediction algorithm

Figure 10 Comparison of prediction effects of three algorithms

Table 6 Comparison of computational efficiency of fouralgorithms

Algorithm Training timeaverage (sec)

Prediction timeaverage (sec)

Parallel LSTM 436374 1002Single LSTM 1218337 3662BPNN 1094911 2253ARMIA 54696 215

Complexity 11

Renewable and Sustainable Energy Reviews vol 54 pp 396ndash405 2016

[3] C Tu X He Z Shuai and F Jiang ldquoBig data issues in smartgrid-a reviewrdquo Renewable and Sustainable Energy Reviewsvol 79 pp 1099ndash1107 2017

[4] Y Zhang R Yang K Zhang H Jiang and J J ZhangldquoConsumption behavior analytics-aided energy forecastingand dispatchrdquo IEEE Intelligent Systems vol 32 no 4pp 59ndash63 2017

[5] J Deng F Wang Y Chen and X Zhao ldquoFrom industry 40 toenergy 50 the concept connotation and system framework ofintelligent energy systemsrdquo Journal of Automation vol 41no 12 pp 2003ndash2016 2015

[6] Q Sun L Shi Y Ni D Si and J Zhu ldquoAn enhanced cas-cading failure model integrating data mining techniquerdquoProtection and Control of Modern Power Systems vol 2 2017

[7] D Zheng A T Eseye J Zhang and H Li ldquoShort-term windpower forecasting using a double-stage hierarchical ANFISapproach for energy management in microgridsrdquo Protectionand Control of Modern Power Systems vol 2 pp 1ndash10 2017

[8] Z Li L Ye Y Zhao X Song J Teng and J Jin ldquoShort-termwind power prediction based on extreme learning machinewith error correctionrdquo Protection and Control of ModernPower Systems vol 1 no 1 pp 1ndash8 2016

[9] H Z Wang Z X Lei X Zhang B Zhou and J C Peng ldquoAreview of deep learning for renewable energy forecastingrdquoEnergy Conversion and Management vol 198 pp 1ndash16 2019

[10] H Z Wang G B Wang G Q Li J C Peng and Y T LiuldquoDeep belief network based deterministic and probabilisticwind speed forecasting approachrdquo Applied Energy vol 182pp 80ndash93 2016

[11] M-R Chen G-Q Zeng K-D Lu and J Weng ldquoA two-layernonlinear combination method for short-term wind speedprediction based on ELM ENN and LSTMrdquo IEEE Internet ofings Journal vol 6 no 4 pp 6997ndash7010 2019

[12] H Liu X Mi and Y Li ldquoSmart multi-step deep learningmodel for wind speed forecasting based on variational modedecomposition singular spectrum analysis LSTM networkand ELMrdquo Energy Conversion and Management vol 159pp 54ndash64 2018

[13] J Chen G-Q Zeng W Zhou W Du and K-D Lu ldquoWindspeed forecasting using nonlinear-learning ensemble of deeplearning time series prediction and extremal optimizationrdquoEnergy Conversion and Management vol 165 pp 681ndash6952018

[14] W Kong Z Y Dong Y Jia D J Hill Y Xu and Y ZhangldquoShort-term residential load forecasting based on LSTM re-current neural networkrdquo IEEE Transactions on Smart Gridvol 10 no 1 2017

[15] H Gao W Huang X Yang Y Duan and Y Yin ldquoTowardservice selection for workflow reconfiguration an interface-based computing solutionrdquo Future Generation ComputerSystems vol 87 pp 298ndash311 2018

[16] L Qi Q He F Chen et al ldquoFinding all you need web APIsrecommendation in web of things through keywords searchrdquoIEEE Transactions on Computational Social Systems vol 6no 5 2019

[17] L Qi R Wang C Hu S Li Q He and X Xu ldquoTime-awaredistributed service recommendation with privacy-preserva-tionrdquo Information Sciences vol 480 pp 354ndash364 2019

[18] Y Zhang and Q Yang ldquoAn overview of multi-task learningrdquoNational Science Review vol 5 no 1 pp 34ndash47 2018

[19] A Argyriou M T Evgeniou and M Pontil ldquoConvex multi-task feature learningrdquo Machine Learning vol 73 no 3pp 243ndash272 2008

[20] T Jebara ldquoMultitask sparsity via maximum entropy dis-criminationrdquo Journal of Machine Learning Research vol 12no 1 pp 75ndash110 2011

[21] L Qi Y Chen Y Yuan S Fu X Zhang and X Xu ldquoA QoS-aware virtual machine scheduling method for energy con-servation in cloud-based cyber-physical systemsrdquo WorldWide Web vol 5 pp 1ndash23 2019

[22] R Zhang P Xie C Wang G Liu and S Wan ldquoClassifyingtransportation mode and speed from trajectory data via deepmulti-scale learningrdquo Computer Networks vol 162 p 1068612019

[23] W Wang D Wang H Jia et al ldquoReview of steady-stateanalysis of typical regional integrated energy system under thebackground of energy internetrdquo Proceedings of the CSEE12pp 3292ndash3306 2016

[24] S Ding S Qu Y Xi and S Wan ldquoStimulus-driven andconcept-driven analysis for image caption generationrdquoNeurocomputing 2019

[25] P Vrba V Marik P Siano et al ldquoA review of agent andservice-oriented concepts applied to intelligent energy sys-temsrdquo IEEE Transactions on Industrial Informatics vol 10no 3 pp 1890ndash1903 2014

[26] W Jia C Kang and Q Chen ldquoAnalysis on demand-sideinteractive response capability for power system dispatch in asmart grid frameworkrdquo Electric Power Systems Researchvol 90 pp 11ndash17 2012

[27] Z C Lipton J Berkowitz and C Elkan ldquoA critical review ofrecurrent neural networks for sequence learningrdquo CoRR2015 httpsarxivorgpdf150600019pdf

[28] T N Sainath O Vinyals A Senior and H Sak ldquoCon-volutional long short-term memory fully connected deepneural networksrdquo in Proceedings of the 2015 IEEE In-ternational Conference on Acoustics Speech and Signal Pro-cessing (ICASSP) pp 4580ndash4584 IEEE Brisbane AustraliaApril 2015

[29] Y Guo Z Cheng L Nie YWang J Ma andM KankanhallildquoAttentive long short-term preference modeling for person-alized product searchrdquo ACM Transactions on InformationSystems (TOIS) vol 37 no 2 p 19 2019

[30] A Graves ldquoLong short-term memoryrdquo Neural Computationvol 9 pp 1735ndash1780 1997

[31] F A Gers J Schmidhuber and F Cummins ldquoLearning toforget continual prediction with LSTMrdquo Neural Computa-tion vol 12 no 10 pp 2451ndash2471 2000

[32] M L Seltzer and J Droppo ldquoMulti-task learning in deepneural networks for improved phoneme recognitionrdquo inProceedings of the 2013 IEEE International Conference onAcoustics Speech and Signal Processing pp 6965ndash6969 IEEEVancouver Canada May 2013

[33] S Wan Y Zhao T Wang Z Gu Q H Abbasi andK-K R Choo ldquoMulti-dimensional data indexing and rangequery processing via Voronoi diagram for internet of thingsrdquoFuture Generation Computer Systems vol 91 pp 382ndash3912019

[34] Z Gao D Y Wang S H Wan H Zhang and Y L WangldquoCognitive-inspired class-statistic matching with triple-con-strain for camera free 3D object retrievalrdquo Future GenerationComputer Systems vol 94 pp 641ndash653 2019

[35] C A Kang A R Brandt and L J Durlofsky ldquoOptimaloperation of an integrated energy system including fossil fuel

12 Complexity

power generation CO2 capture and windrdquo Energy vol 36no 12 pp 6806ndash6820 2011

[36] Q Zhang SWan BWang DW Gao and HMa ldquoAnomalydetection based on randommatrix theory for industrial powersystemsrdquo Journal of Systems Architecture vol 95 pp 67ndash742019

[37] Z Gao H Z Xuan H Zhang S Wan and K K R ChooldquoAdaptive fusion and category-level dictionary learningmodel for multi-view human action recognitionrdquo IEEE In-ternet of ings Journal p 1 2019

[38] WWu K Chen Y Qiao and Z Lu ldquoProbabilistic short-termwind power forecasting based on deep neural networksrdquo inProceedings of the 2016 International Conference on Proba-bilistic Methods Applied to Power Systems (PMAPS) pp 1ndash8Beijing China October 2016

[39] K M Powell A Sriprasad W J Cole and T F EdgarldquoHeating cooling and electrical load forecasting for a large-scale district energy systemrdquo Energy vol 74 pp 877ndash8852014

[40] A Peimankar S J Weddell T Jalal and A C LapthornldquoMulti-objective ensemble forecasting with an application topower transformersrdquo Applied Soft Computing vol 68pp 233ndash248 2018

[41] G-Q Zeng J Chen L-M Li et al ldquoAn improved multi-objective population-based extremal optimization algorithmwith polynomial mutationrdquo Information Sciences vol 330pp 49ndash73 2016

[42] Y-L Hu and L Chen ldquoA nonlinear hybrid wind speedforecasting model using LSTM network hysteretic ELM andDifferential Evolution algorithmrdquo Energy Conversion andManagement vol 173 pp 123ndash142 2018

[43] G-Q Zeng J Chen Y-X Dai L-M Li C-W Zheng andM-R Chen ldquoAn improved Design of fractional order PIDcontroller for automatic regulator voltage system based onmulti-objective extremal optimizationrdquo Neurocomputingvol 160 pp 173ndash184 2015

[44] X Xu R Gu F Dai L Qi and S Wan ldquoMulti-objectivecomputation offloading for internet of vehicles in cloud-edgecomputingrdquo Wireless Networks pp 1ndash19 2019

[45] S Wan X Li Y Xue W Lin and X Xu ldquoEfficient com-putation offloading for internet of vehicles in edge comput-ing-assisted 5G networksrdquo e Journal of Supercomputingpp 1ndash30 2019

[46] B Wang S Wan X Zhang and K-K R Choo ldquoA novelindex for assessing the robustness of integrated electricalnetwork and a natural gas networkrdquo IEEE Access vol 6pp 40400ndash40410 2018

[47] Y Yin F Yu Y Xu L Yu and J Mu ldquoNetwork location-aware service recommendation with random walk in cyber-physical systemsrdquo Sensors vol 17 no 9 p 2059 2017

Complexity 13

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Page 9: Parallel LSTM-Based Regional Integrated Energy System ...downloads.hindawi.com/journals/complexity/2019/7414318.pdf · Characteristics of Regional Integrated Energy System “Information

should increase as a whole +is is due to the cold weather inthe rainy scenes which affects the userrsquos energy use behaviorand increases the userrsquos demand for electricity and heat thusincreasing the electrical load and heat load

621 Comparative Analysis of Multienergy Load InformationInteraction In order to verify the effect of interactivelearning of horizontal and vertical multienergy informationthe parallel LSTM predictions considering user energy be-havior and parallel LSTM prediction without consideringuser energy behavior and the results of single LSTM pre-diction without considering user energy behavior are pro-posed to conduct a comparative analysis In order to ensurethe fairness of the experimental test this paper uses the samedataset for training and testing

+e comparison experiment setup is shown in Table 3 Inthe parallel LSTM learning without considering the userrsquosenergy behavior the input sample is based on the originalprediction model A and the lighting load data reflecting theuser behavior in the input sample are not considered +eprediction is based only on the historical data of the pre-dicted object the model C does not consider the singleLSTMprediction of the userrsquos energy behavior and the inputand output of each subprediction model only consider theinput and output of a single energy and there is also noparameter sharing between the LSTM unit of the different

submodels Model B uses parallel LSTM learning +ere isparameter sharing among each subtask which belongs to amultitask learning Model C uses a single LSTM learningand each subtask does not affect each other It belongs to asingle-task learning

+e prediction accuracy of the multienergy charge in-formation interaction comparison experiment is shown inTable 4 and Figure 9 +rough the analysis of Table 4 andFigure 9 it shows that the average prediction accuracy of theweights of the models A B and C are 09009 08917 and08728 respectively +e prediction result of the model A isthe best the model B is the second and the model C is thethird +at is considering the interaction of vertical andhorizontal information the energy predictions on both sidesof the source and the load show better prediction accuracy

Comparing the prediction results of Model B and ModelC for the five subprediction tasks when the different qualityenergies are predicted together in parallel a higher precisionthan the single prediction is obtained It can be seen thatparallel LSTM learning considering horizontal multienergyinformation interaction utilizes the coupled data betweendifferent systems to fully exploit the interactive informationbetween different quality energy sources of the integratedenergy system so that the prediction performance of acertain energy can be improved In addition considering themultienergy horizontal information interaction the pre-dicted accuracy of the load-side electrical load heat load and

True valuePredictive value

5 10 15 20 250t (h)

100

150

200

250

300

350W

ind

pow

er (k

W)

(a)

True valuePredictive value

5 10 15 20 250t (h)

0

20

40

60

80

100

120

PV p

ower

(kW

)(b)

True valuePredictive value

5 10 15 20 250t (h)

300

400

500

600

700

Elec

tric

load

(kW

)

(c)

True valuePredictive value

5 10 15 20 250t (h)

60

80

100

120

140

160

Hea

t loa

d (k

W)

(d)

True valuePredictive value

5 10 15 20 250t (h)

0

100

200

300

400

500

Col

d lo

ad (k

W)

(e)

Figure 8+e results of source-load bilateral energy prediction within a rainy day (a) Prediction of wind power (b) Prediction of PV power(c) Prediction of electric load (d) Prediction of heat load (e) Prediction of cold load

Complexity 9

cold load are increased by 313 253 and 267 re-spectively while the predicted accuracy of the source-sidewind power and photovoltaic is increased by 152 and146 and the energy prediction accuracy of the charge sideis relatively high It can be seen that the source-side un-certainty resources are also affected by other energy in-teractions in the system but the impact is small while thecharge side demand is largely affected by other energy in-teractions in the system And the information interaction ismore intimate between each other

Comparing the prediction results of Model A and ModelB when the userrsquos energy behavior information interactsbetween the source and the load in vertical direction theaverage weight accuracy of the overall prediction is increasedby 103 Comparing the performance improvement effectsof the five subprediction tasks it can be found that thepercentage improvement of PV prediction accuracy is185 and the improvement effect is relatively best fol-lowed by the electrical load while the prediction accuracy ofwind power heat load and cooling load is not obvious Itcan be seen that considering the source-load vertical in-formation interaction can improve the energy predictionaccuracy on both sides of the source and the load but thespecific user energy behavior has different differences in theenergy prediction of different types of users and the user

lighting behavior is obviously related to photovoltaic powergeneration and user electricity +e interaction betweenloads is more tight and its effect on PV prediction andelectric load forecasting is more obvious

622 Comparison of Deep Neural Network Algorithms withOther Algorithms In order to verify the prediction effect ofthe LSTM deep neural network algorithm the LSTM modeland the autoregressive integrated moving average model(ARIMA) algorithm and the classic BP neural network(BPNN) algorithm in this paper are proposed +e simu-lation results of them were compared and analyzed In orderto ensure the fairness of the experimental test this paper usesthe same dataset for training and testing and both use asingle-task prediction strategy +e prediction results of thethree algorithms are shown in Table 5 and Figure 10

Observing the prediction results the prediction effect ofARIMA algorithm is the most general and the predictionaccuracy of BPNN algorithm is much higher than that ofARIMA algorithm +e prediction accuracy of each task ofLSTM deep learning prediction algorithm proposed in thispaper is the highest +e ARIMA algorithm can not effec-tively use the historical information in the long-term se-quence because it only uses a part of the historicalinformation to train which leads to the large predictionerror +e LSTM algorithm can fully utilize its strongnonlinear learning ability and long-term and short-termmemory advantages to exploit the information value of long-term historical data so the grasp of the energy prediction lawis more precise

In addition comparing the energy prediction results ofthe LSTM algorithm and the BPNN algorithm one by onethe LSTM algorithm predicts that the accuracy of thethermal load and the cold load prediction is more obvious Itcan be seen that due to the greater inertia of the thermalsystem the time scale of the change law is longer whichmakes the prediction accuracy of the long-term and short-term deep learning on the heat load and the cold loadsignificantly higher than the electric load

In order to evaluate the computational efficiency ofalgorithms four forecasting scenarios are performed tentimes so that the forecasting results can reliably be sum-marized as in Table 6 Parallel LSTM and Single LSTM havelonger training time but the prediction time is kept within

Table 3 Contrast experiment setting of multisource source information interaction

Model Vertical interaction Horizontal interactionModel A User energy behavior is considered Parallel LSTM learningModel B User energy behavior is not considered Parallel LSTM learningModel C User energy behavior is not considered Single-LSTM learning

Table 4 +ree kinds of model prediction results of information interaction comparison experiment

Precision Wind power PV power Electric load Heat load Cold load Mean precisionModel A 08602 08735 09236 09445 09407 09009Model B 08542 08576 09113 09401 09368 08917Model C 08414 08453 08836 09169 09124 08728

Model AModel BModel C

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

PV Electric Heat Cold MeanWindPrediction task

Figure 9 Comparison of prediction effects of three models

10 Complexity

one minute and can meet the requirements of short-termforecasting Further analysis shows that the parallel trainingtime and prediction time of parallel LSTM is greatly im-proved compared to single LSTM due to the combination offive subtasks into one multitask learning In summary al-though the LSTM algorithm is time consuming in trainingtime for short-term prediction the training process doesnot require online completion and the prediction time ofthe LSTM algorithm especially parallel LSTM can fullymeet the requirements of short-term prediction Consid-ering the prediction accuracy the sacrifice of the LSTMalgorithm in computing time is worthwhile

7 Conclusion

In this paper the horizontal and vertical information interactioncharacteristics between multiple energy loads in RIES areconsidered Using LSTM network deep learning algorithm a

parallel LSTM-based multienergy-load information interactiveshort-term energy prediction method is proposed+rough thesimulation of the test data the following conclusions are drawn

(1) +e ldquoinformation interactionrdquo strategy proposed inthis paper can effectively improve the predictionperformance of source-load energy prediction inregional integrated energy systems

(2) Parallel prediction considering horizontal in-teraction mainly affects the prediction effect on theload side while the information sharing of user il-lumination behavior considering vertical interactionmainly affects the improvement of PV predictioneffect

(3) Compared with the conventional prediction algo-rithm the LSTM deep learning algorithm can learnthe deep hidden information of historical dataimprove the prediction accuracy while meetingcalculation time and prove the effectiveness of thealgorithm in the prediction of time series featuresequence

However the performance of the proposed parallelLSTM algorithm may be further enhanced by some evo-lutionary algorithms or multiobjective optimization algo-rithms such as differential evolution algorithm [40]multiobjective extremal optimization [41] and evolutionarymultiobjective optimization algorithms [42ndash47] Of coursesome significant topics are worth researching on the opti-mization of the proposed parallel LSTM by multiobjectiveoptimization

Data Availability

+e data used to support the findings of this study areavailable from the corresponding author upon request

Conflicts of Interest

+e authors declare no potential conflicts of interest

Acknowledgments

+is work was fully supported by the National Key Researchand Development Program of China (no 2018YFB0904200)

References

[1] A Q Huang M L Crow G T Heydt J P Zheng andS J Dale ldquo+e future renewable electric energy delivery andmanagement (FREEDM) system the energy internetrdquo Pro-ceedings of the IEEE vol 99 no 1 pp 133ndash148 2011

[2] I Colak S Sagiroglu G Fulli M Yesilbudak and C CovrigldquoA survey on the critical issues in smart grid technologiesrdquo

Table 5 Comparison of prediction results of three algorithms

Precision Wind power PV power Electric load Heat load Cold load Mean precisionLSTM 08414 08453 08836 09169 09124 08728BPNN 08216 08314 08695 08912 08889 08542ARMIA 07733 07824 08282 08454 08408 08075

Prediction precision ofwind powerPrediction precision ofPV powerPrediction precision ofelectric load

Prediction precision ofhead loadPrediction precision ofcold loadWeight average predictionprecision

07

075

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

BPNN ARMIALSTMPrediction algorithm

Figure 10 Comparison of prediction effects of three algorithms

Table 6 Comparison of computational efficiency of fouralgorithms

Algorithm Training timeaverage (sec)

Prediction timeaverage (sec)

Parallel LSTM 436374 1002Single LSTM 1218337 3662BPNN 1094911 2253ARMIA 54696 215

Complexity 11

Renewable and Sustainable Energy Reviews vol 54 pp 396ndash405 2016

[3] C Tu X He Z Shuai and F Jiang ldquoBig data issues in smartgrid-a reviewrdquo Renewable and Sustainable Energy Reviewsvol 79 pp 1099ndash1107 2017

[4] Y Zhang R Yang K Zhang H Jiang and J J ZhangldquoConsumption behavior analytics-aided energy forecastingand dispatchrdquo IEEE Intelligent Systems vol 32 no 4pp 59ndash63 2017

[5] J Deng F Wang Y Chen and X Zhao ldquoFrom industry 40 toenergy 50 the concept connotation and system framework ofintelligent energy systemsrdquo Journal of Automation vol 41no 12 pp 2003ndash2016 2015

[6] Q Sun L Shi Y Ni D Si and J Zhu ldquoAn enhanced cas-cading failure model integrating data mining techniquerdquoProtection and Control of Modern Power Systems vol 2 2017

[7] D Zheng A T Eseye J Zhang and H Li ldquoShort-term windpower forecasting using a double-stage hierarchical ANFISapproach for energy management in microgridsrdquo Protectionand Control of Modern Power Systems vol 2 pp 1ndash10 2017

[8] Z Li L Ye Y Zhao X Song J Teng and J Jin ldquoShort-termwind power prediction based on extreme learning machinewith error correctionrdquo Protection and Control of ModernPower Systems vol 1 no 1 pp 1ndash8 2016

[9] H Z Wang Z X Lei X Zhang B Zhou and J C Peng ldquoAreview of deep learning for renewable energy forecastingrdquoEnergy Conversion and Management vol 198 pp 1ndash16 2019

[10] H Z Wang G B Wang G Q Li J C Peng and Y T LiuldquoDeep belief network based deterministic and probabilisticwind speed forecasting approachrdquo Applied Energy vol 182pp 80ndash93 2016

[11] M-R Chen G-Q Zeng K-D Lu and J Weng ldquoA two-layernonlinear combination method for short-term wind speedprediction based on ELM ENN and LSTMrdquo IEEE Internet ofings Journal vol 6 no 4 pp 6997ndash7010 2019

[12] H Liu X Mi and Y Li ldquoSmart multi-step deep learningmodel for wind speed forecasting based on variational modedecomposition singular spectrum analysis LSTM networkand ELMrdquo Energy Conversion and Management vol 159pp 54ndash64 2018

[13] J Chen G-Q Zeng W Zhou W Du and K-D Lu ldquoWindspeed forecasting using nonlinear-learning ensemble of deeplearning time series prediction and extremal optimizationrdquoEnergy Conversion and Management vol 165 pp 681ndash6952018

[14] W Kong Z Y Dong Y Jia D J Hill Y Xu and Y ZhangldquoShort-term residential load forecasting based on LSTM re-current neural networkrdquo IEEE Transactions on Smart Gridvol 10 no 1 2017

[15] H Gao W Huang X Yang Y Duan and Y Yin ldquoTowardservice selection for workflow reconfiguration an interface-based computing solutionrdquo Future Generation ComputerSystems vol 87 pp 298ndash311 2018

[16] L Qi Q He F Chen et al ldquoFinding all you need web APIsrecommendation in web of things through keywords searchrdquoIEEE Transactions on Computational Social Systems vol 6no 5 2019

[17] L Qi R Wang C Hu S Li Q He and X Xu ldquoTime-awaredistributed service recommendation with privacy-preserva-tionrdquo Information Sciences vol 480 pp 354ndash364 2019

[18] Y Zhang and Q Yang ldquoAn overview of multi-task learningrdquoNational Science Review vol 5 no 1 pp 34ndash47 2018

[19] A Argyriou M T Evgeniou and M Pontil ldquoConvex multi-task feature learningrdquo Machine Learning vol 73 no 3pp 243ndash272 2008

[20] T Jebara ldquoMultitask sparsity via maximum entropy dis-criminationrdquo Journal of Machine Learning Research vol 12no 1 pp 75ndash110 2011

[21] L Qi Y Chen Y Yuan S Fu X Zhang and X Xu ldquoA QoS-aware virtual machine scheduling method for energy con-servation in cloud-based cyber-physical systemsrdquo WorldWide Web vol 5 pp 1ndash23 2019

[22] R Zhang P Xie C Wang G Liu and S Wan ldquoClassifyingtransportation mode and speed from trajectory data via deepmulti-scale learningrdquo Computer Networks vol 162 p 1068612019

[23] W Wang D Wang H Jia et al ldquoReview of steady-stateanalysis of typical regional integrated energy system under thebackground of energy internetrdquo Proceedings of the CSEE12pp 3292ndash3306 2016

[24] S Ding S Qu Y Xi and S Wan ldquoStimulus-driven andconcept-driven analysis for image caption generationrdquoNeurocomputing 2019

[25] P Vrba V Marik P Siano et al ldquoA review of agent andservice-oriented concepts applied to intelligent energy sys-temsrdquo IEEE Transactions on Industrial Informatics vol 10no 3 pp 1890ndash1903 2014

[26] W Jia C Kang and Q Chen ldquoAnalysis on demand-sideinteractive response capability for power system dispatch in asmart grid frameworkrdquo Electric Power Systems Researchvol 90 pp 11ndash17 2012

[27] Z C Lipton J Berkowitz and C Elkan ldquoA critical review ofrecurrent neural networks for sequence learningrdquo CoRR2015 httpsarxivorgpdf150600019pdf

[28] T N Sainath O Vinyals A Senior and H Sak ldquoCon-volutional long short-term memory fully connected deepneural networksrdquo in Proceedings of the 2015 IEEE In-ternational Conference on Acoustics Speech and Signal Pro-cessing (ICASSP) pp 4580ndash4584 IEEE Brisbane AustraliaApril 2015

[29] Y Guo Z Cheng L Nie YWang J Ma andM KankanhallildquoAttentive long short-term preference modeling for person-alized product searchrdquo ACM Transactions on InformationSystems (TOIS) vol 37 no 2 p 19 2019

[30] A Graves ldquoLong short-term memoryrdquo Neural Computationvol 9 pp 1735ndash1780 1997

[31] F A Gers J Schmidhuber and F Cummins ldquoLearning toforget continual prediction with LSTMrdquo Neural Computa-tion vol 12 no 10 pp 2451ndash2471 2000

[32] M L Seltzer and J Droppo ldquoMulti-task learning in deepneural networks for improved phoneme recognitionrdquo inProceedings of the 2013 IEEE International Conference onAcoustics Speech and Signal Processing pp 6965ndash6969 IEEEVancouver Canada May 2013

[33] S Wan Y Zhao T Wang Z Gu Q H Abbasi andK-K R Choo ldquoMulti-dimensional data indexing and rangequery processing via Voronoi diagram for internet of thingsrdquoFuture Generation Computer Systems vol 91 pp 382ndash3912019

[34] Z Gao D Y Wang S H Wan H Zhang and Y L WangldquoCognitive-inspired class-statistic matching with triple-con-strain for camera free 3D object retrievalrdquo Future GenerationComputer Systems vol 94 pp 641ndash653 2019

[35] C A Kang A R Brandt and L J Durlofsky ldquoOptimaloperation of an integrated energy system including fossil fuel

12 Complexity

power generation CO2 capture and windrdquo Energy vol 36no 12 pp 6806ndash6820 2011

[36] Q Zhang SWan BWang DW Gao and HMa ldquoAnomalydetection based on randommatrix theory for industrial powersystemsrdquo Journal of Systems Architecture vol 95 pp 67ndash742019

[37] Z Gao H Z Xuan H Zhang S Wan and K K R ChooldquoAdaptive fusion and category-level dictionary learningmodel for multi-view human action recognitionrdquo IEEE In-ternet of ings Journal p 1 2019

[38] WWu K Chen Y Qiao and Z Lu ldquoProbabilistic short-termwind power forecasting based on deep neural networksrdquo inProceedings of the 2016 International Conference on Proba-bilistic Methods Applied to Power Systems (PMAPS) pp 1ndash8Beijing China October 2016

[39] K M Powell A Sriprasad W J Cole and T F EdgarldquoHeating cooling and electrical load forecasting for a large-scale district energy systemrdquo Energy vol 74 pp 877ndash8852014

[40] A Peimankar S J Weddell T Jalal and A C LapthornldquoMulti-objective ensemble forecasting with an application topower transformersrdquo Applied Soft Computing vol 68pp 233ndash248 2018

[41] G-Q Zeng J Chen L-M Li et al ldquoAn improved multi-objective population-based extremal optimization algorithmwith polynomial mutationrdquo Information Sciences vol 330pp 49ndash73 2016

[42] Y-L Hu and L Chen ldquoA nonlinear hybrid wind speedforecasting model using LSTM network hysteretic ELM andDifferential Evolution algorithmrdquo Energy Conversion andManagement vol 173 pp 123ndash142 2018

[43] G-Q Zeng J Chen Y-X Dai L-M Li C-W Zheng andM-R Chen ldquoAn improved Design of fractional order PIDcontroller for automatic regulator voltage system based onmulti-objective extremal optimizationrdquo Neurocomputingvol 160 pp 173ndash184 2015

[44] X Xu R Gu F Dai L Qi and S Wan ldquoMulti-objectivecomputation offloading for internet of vehicles in cloud-edgecomputingrdquo Wireless Networks pp 1ndash19 2019

[45] S Wan X Li Y Xue W Lin and X Xu ldquoEfficient com-putation offloading for internet of vehicles in edge comput-ing-assisted 5G networksrdquo e Journal of Supercomputingpp 1ndash30 2019

[46] B Wang S Wan X Zhang and K-K R Choo ldquoA novelindex for assessing the robustness of integrated electricalnetwork and a natural gas networkrdquo IEEE Access vol 6pp 40400ndash40410 2018

[47] Y Yin F Yu Y Xu L Yu and J Mu ldquoNetwork location-aware service recommendation with random walk in cyber-physical systemsrdquo Sensors vol 17 no 9 p 2059 2017

Complexity 13

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Page 10: Parallel LSTM-Based Regional Integrated Energy System ...downloads.hindawi.com/journals/complexity/2019/7414318.pdf · Characteristics of Regional Integrated Energy System “Information

cold load are increased by 313 253 and 267 re-spectively while the predicted accuracy of the source-sidewind power and photovoltaic is increased by 152 and146 and the energy prediction accuracy of the charge sideis relatively high It can be seen that the source-side un-certainty resources are also affected by other energy in-teractions in the system but the impact is small while thecharge side demand is largely affected by other energy in-teractions in the system And the information interaction ismore intimate between each other

Comparing the prediction results of Model A and ModelB when the userrsquos energy behavior information interactsbetween the source and the load in vertical direction theaverage weight accuracy of the overall prediction is increasedby 103 Comparing the performance improvement effectsof the five subprediction tasks it can be found that thepercentage improvement of PV prediction accuracy is185 and the improvement effect is relatively best fol-lowed by the electrical load while the prediction accuracy ofwind power heat load and cooling load is not obvious Itcan be seen that considering the source-load vertical in-formation interaction can improve the energy predictionaccuracy on both sides of the source and the load but thespecific user energy behavior has different differences in theenergy prediction of different types of users and the user

lighting behavior is obviously related to photovoltaic powergeneration and user electricity +e interaction betweenloads is more tight and its effect on PV prediction andelectric load forecasting is more obvious

622 Comparison of Deep Neural Network Algorithms withOther Algorithms In order to verify the prediction effect ofthe LSTM deep neural network algorithm the LSTM modeland the autoregressive integrated moving average model(ARIMA) algorithm and the classic BP neural network(BPNN) algorithm in this paper are proposed +e simu-lation results of them were compared and analyzed In orderto ensure the fairness of the experimental test this paper usesthe same dataset for training and testing and both use asingle-task prediction strategy +e prediction results of thethree algorithms are shown in Table 5 and Figure 10

Observing the prediction results the prediction effect ofARIMA algorithm is the most general and the predictionaccuracy of BPNN algorithm is much higher than that ofARIMA algorithm +e prediction accuracy of each task ofLSTM deep learning prediction algorithm proposed in thispaper is the highest +e ARIMA algorithm can not effec-tively use the historical information in the long-term se-quence because it only uses a part of the historicalinformation to train which leads to the large predictionerror +e LSTM algorithm can fully utilize its strongnonlinear learning ability and long-term and short-termmemory advantages to exploit the information value of long-term historical data so the grasp of the energy prediction lawis more precise

In addition comparing the energy prediction results ofthe LSTM algorithm and the BPNN algorithm one by onethe LSTM algorithm predicts that the accuracy of thethermal load and the cold load prediction is more obvious Itcan be seen that due to the greater inertia of the thermalsystem the time scale of the change law is longer whichmakes the prediction accuracy of the long-term and short-term deep learning on the heat load and the cold loadsignificantly higher than the electric load

In order to evaluate the computational efficiency ofalgorithms four forecasting scenarios are performed tentimes so that the forecasting results can reliably be sum-marized as in Table 6 Parallel LSTM and Single LSTM havelonger training time but the prediction time is kept within

Table 3 Contrast experiment setting of multisource source information interaction

Model Vertical interaction Horizontal interactionModel A User energy behavior is considered Parallel LSTM learningModel B User energy behavior is not considered Parallel LSTM learningModel C User energy behavior is not considered Single-LSTM learning

Table 4 +ree kinds of model prediction results of information interaction comparison experiment

Precision Wind power PV power Electric load Heat load Cold load Mean precisionModel A 08602 08735 09236 09445 09407 09009Model B 08542 08576 09113 09401 09368 08917Model C 08414 08453 08836 09169 09124 08728

Model AModel BModel C

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

PV Electric Heat Cold MeanWindPrediction task

Figure 9 Comparison of prediction effects of three models

10 Complexity

one minute and can meet the requirements of short-termforecasting Further analysis shows that the parallel trainingtime and prediction time of parallel LSTM is greatly im-proved compared to single LSTM due to the combination offive subtasks into one multitask learning In summary al-though the LSTM algorithm is time consuming in trainingtime for short-term prediction the training process doesnot require online completion and the prediction time ofthe LSTM algorithm especially parallel LSTM can fullymeet the requirements of short-term prediction Consid-ering the prediction accuracy the sacrifice of the LSTMalgorithm in computing time is worthwhile

7 Conclusion

In this paper the horizontal and vertical information interactioncharacteristics between multiple energy loads in RIES areconsidered Using LSTM network deep learning algorithm a

parallel LSTM-based multienergy-load information interactiveshort-term energy prediction method is proposed+rough thesimulation of the test data the following conclusions are drawn

(1) +e ldquoinformation interactionrdquo strategy proposed inthis paper can effectively improve the predictionperformance of source-load energy prediction inregional integrated energy systems

(2) Parallel prediction considering horizontal in-teraction mainly affects the prediction effect on theload side while the information sharing of user il-lumination behavior considering vertical interactionmainly affects the improvement of PV predictioneffect

(3) Compared with the conventional prediction algo-rithm the LSTM deep learning algorithm can learnthe deep hidden information of historical dataimprove the prediction accuracy while meetingcalculation time and prove the effectiveness of thealgorithm in the prediction of time series featuresequence

However the performance of the proposed parallelLSTM algorithm may be further enhanced by some evo-lutionary algorithms or multiobjective optimization algo-rithms such as differential evolution algorithm [40]multiobjective extremal optimization [41] and evolutionarymultiobjective optimization algorithms [42ndash47] Of coursesome significant topics are worth researching on the opti-mization of the proposed parallel LSTM by multiobjectiveoptimization

Data Availability

+e data used to support the findings of this study areavailable from the corresponding author upon request

Conflicts of Interest

+e authors declare no potential conflicts of interest

Acknowledgments

+is work was fully supported by the National Key Researchand Development Program of China (no 2018YFB0904200)

References

[1] A Q Huang M L Crow G T Heydt J P Zheng andS J Dale ldquo+e future renewable electric energy delivery andmanagement (FREEDM) system the energy internetrdquo Pro-ceedings of the IEEE vol 99 no 1 pp 133ndash148 2011

[2] I Colak S Sagiroglu G Fulli M Yesilbudak and C CovrigldquoA survey on the critical issues in smart grid technologiesrdquo

Table 5 Comparison of prediction results of three algorithms

Precision Wind power PV power Electric load Heat load Cold load Mean precisionLSTM 08414 08453 08836 09169 09124 08728BPNN 08216 08314 08695 08912 08889 08542ARMIA 07733 07824 08282 08454 08408 08075

Prediction precision ofwind powerPrediction precision ofPV powerPrediction precision ofelectric load

Prediction precision ofhead loadPrediction precision ofcold loadWeight average predictionprecision

07

075

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

BPNN ARMIALSTMPrediction algorithm

Figure 10 Comparison of prediction effects of three algorithms

Table 6 Comparison of computational efficiency of fouralgorithms

Algorithm Training timeaverage (sec)

Prediction timeaverage (sec)

Parallel LSTM 436374 1002Single LSTM 1218337 3662BPNN 1094911 2253ARMIA 54696 215

Complexity 11

Renewable and Sustainable Energy Reviews vol 54 pp 396ndash405 2016

[3] C Tu X He Z Shuai and F Jiang ldquoBig data issues in smartgrid-a reviewrdquo Renewable and Sustainable Energy Reviewsvol 79 pp 1099ndash1107 2017

[4] Y Zhang R Yang K Zhang H Jiang and J J ZhangldquoConsumption behavior analytics-aided energy forecastingand dispatchrdquo IEEE Intelligent Systems vol 32 no 4pp 59ndash63 2017

[5] J Deng F Wang Y Chen and X Zhao ldquoFrom industry 40 toenergy 50 the concept connotation and system framework ofintelligent energy systemsrdquo Journal of Automation vol 41no 12 pp 2003ndash2016 2015

[6] Q Sun L Shi Y Ni D Si and J Zhu ldquoAn enhanced cas-cading failure model integrating data mining techniquerdquoProtection and Control of Modern Power Systems vol 2 2017

[7] D Zheng A T Eseye J Zhang and H Li ldquoShort-term windpower forecasting using a double-stage hierarchical ANFISapproach for energy management in microgridsrdquo Protectionand Control of Modern Power Systems vol 2 pp 1ndash10 2017

[8] Z Li L Ye Y Zhao X Song J Teng and J Jin ldquoShort-termwind power prediction based on extreme learning machinewith error correctionrdquo Protection and Control of ModernPower Systems vol 1 no 1 pp 1ndash8 2016

[9] H Z Wang Z X Lei X Zhang B Zhou and J C Peng ldquoAreview of deep learning for renewable energy forecastingrdquoEnergy Conversion and Management vol 198 pp 1ndash16 2019

[10] H Z Wang G B Wang G Q Li J C Peng and Y T LiuldquoDeep belief network based deterministic and probabilisticwind speed forecasting approachrdquo Applied Energy vol 182pp 80ndash93 2016

[11] M-R Chen G-Q Zeng K-D Lu and J Weng ldquoA two-layernonlinear combination method for short-term wind speedprediction based on ELM ENN and LSTMrdquo IEEE Internet ofings Journal vol 6 no 4 pp 6997ndash7010 2019

[12] H Liu X Mi and Y Li ldquoSmart multi-step deep learningmodel for wind speed forecasting based on variational modedecomposition singular spectrum analysis LSTM networkand ELMrdquo Energy Conversion and Management vol 159pp 54ndash64 2018

[13] J Chen G-Q Zeng W Zhou W Du and K-D Lu ldquoWindspeed forecasting using nonlinear-learning ensemble of deeplearning time series prediction and extremal optimizationrdquoEnergy Conversion and Management vol 165 pp 681ndash6952018

[14] W Kong Z Y Dong Y Jia D J Hill Y Xu and Y ZhangldquoShort-term residential load forecasting based on LSTM re-current neural networkrdquo IEEE Transactions on Smart Gridvol 10 no 1 2017

[15] H Gao W Huang X Yang Y Duan and Y Yin ldquoTowardservice selection for workflow reconfiguration an interface-based computing solutionrdquo Future Generation ComputerSystems vol 87 pp 298ndash311 2018

[16] L Qi Q He F Chen et al ldquoFinding all you need web APIsrecommendation in web of things through keywords searchrdquoIEEE Transactions on Computational Social Systems vol 6no 5 2019

[17] L Qi R Wang C Hu S Li Q He and X Xu ldquoTime-awaredistributed service recommendation with privacy-preserva-tionrdquo Information Sciences vol 480 pp 354ndash364 2019

[18] Y Zhang and Q Yang ldquoAn overview of multi-task learningrdquoNational Science Review vol 5 no 1 pp 34ndash47 2018

[19] A Argyriou M T Evgeniou and M Pontil ldquoConvex multi-task feature learningrdquo Machine Learning vol 73 no 3pp 243ndash272 2008

[20] T Jebara ldquoMultitask sparsity via maximum entropy dis-criminationrdquo Journal of Machine Learning Research vol 12no 1 pp 75ndash110 2011

[21] L Qi Y Chen Y Yuan S Fu X Zhang and X Xu ldquoA QoS-aware virtual machine scheduling method for energy con-servation in cloud-based cyber-physical systemsrdquo WorldWide Web vol 5 pp 1ndash23 2019

[22] R Zhang P Xie C Wang G Liu and S Wan ldquoClassifyingtransportation mode and speed from trajectory data via deepmulti-scale learningrdquo Computer Networks vol 162 p 1068612019

[23] W Wang D Wang H Jia et al ldquoReview of steady-stateanalysis of typical regional integrated energy system under thebackground of energy internetrdquo Proceedings of the CSEE12pp 3292ndash3306 2016

[24] S Ding S Qu Y Xi and S Wan ldquoStimulus-driven andconcept-driven analysis for image caption generationrdquoNeurocomputing 2019

[25] P Vrba V Marik P Siano et al ldquoA review of agent andservice-oriented concepts applied to intelligent energy sys-temsrdquo IEEE Transactions on Industrial Informatics vol 10no 3 pp 1890ndash1903 2014

[26] W Jia C Kang and Q Chen ldquoAnalysis on demand-sideinteractive response capability for power system dispatch in asmart grid frameworkrdquo Electric Power Systems Researchvol 90 pp 11ndash17 2012

[27] Z C Lipton J Berkowitz and C Elkan ldquoA critical review ofrecurrent neural networks for sequence learningrdquo CoRR2015 httpsarxivorgpdf150600019pdf

[28] T N Sainath O Vinyals A Senior and H Sak ldquoCon-volutional long short-term memory fully connected deepneural networksrdquo in Proceedings of the 2015 IEEE In-ternational Conference on Acoustics Speech and Signal Pro-cessing (ICASSP) pp 4580ndash4584 IEEE Brisbane AustraliaApril 2015

[29] Y Guo Z Cheng L Nie YWang J Ma andM KankanhallildquoAttentive long short-term preference modeling for person-alized product searchrdquo ACM Transactions on InformationSystems (TOIS) vol 37 no 2 p 19 2019

[30] A Graves ldquoLong short-term memoryrdquo Neural Computationvol 9 pp 1735ndash1780 1997

[31] F A Gers J Schmidhuber and F Cummins ldquoLearning toforget continual prediction with LSTMrdquo Neural Computa-tion vol 12 no 10 pp 2451ndash2471 2000

[32] M L Seltzer and J Droppo ldquoMulti-task learning in deepneural networks for improved phoneme recognitionrdquo inProceedings of the 2013 IEEE International Conference onAcoustics Speech and Signal Processing pp 6965ndash6969 IEEEVancouver Canada May 2013

[33] S Wan Y Zhao T Wang Z Gu Q H Abbasi andK-K R Choo ldquoMulti-dimensional data indexing and rangequery processing via Voronoi diagram for internet of thingsrdquoFuture Generation Computer Systems vol 91 pp 382ndash3912019

[34] Z Gao D Y Wang S H Wan H Zhang and Y L WangldquoCognitive-inspired class-statistic matching with triple-con-strain for camera free 3D object retrievalrdquo Future GenerationComputer Systems vol 94 pp 641ndash653 2019

[35] C A Kang A R Brandt and L J Durlofsky ldquoOptimaloperation of an integrated energy system including fossil fuel

12 Complexity

power generation CO2 capture and windrdquo Energy vol 36no 12 pp 6806ndash6820 2011

[36] Q Zhang SWan BWang DW Gao and HMa ldquoAnomalydetection based on randommatrix theory for industrial powersystemsrdquo Journal of Systems Architecture vol 95 pp 67ndash742019

[37] Z Gao H Z Xuan H Zhang S Wan and K K R ChooldquoAdaptive fusion and category-level dictionary learningmodel for multi-view human action recognitionrdquo IEEE In-ternet of ings Journal p 1 2019

[38] WWu K Chen Y Qiao and Z Lu ldquoProbabilistic short-termwind power forecasting based on deep neural networksrdquo inProceedings of the 2016 International Conference on Proba-bilistic Methods Applied to Power Systems (PMAPS) pp 1ndash8Beijing China October 2016

[39] K M Powell A Sriprasad W J Cole and T F EdgarldquoHeating cooling and electrical load forecasting for a large-scale district energy systemrdquo Energy vol 74 pp 877ndash8852014

[40] A Peimankar S J Weddell T Jalal and A C LapthornldquoMulti-objective ensemble forecasting with an application topower transformersrdquo Applied Soft Computing vol 68pp 233ndash248 2018

[41] G-Q Zeng J Chen L-M Li et al ldquoAn improved multi-objective population-based extremal optimization algorithmwith polynomial mutationrdquo Information Sciences vol 330pp 49ndash73 2016

[42] Y-L Hu and L Chen ldquoA nonlinear hybrid wind speedforecasting model using LSTM network hysteretic ELM andDifferential Evolution algorithmrdquo Energy Conversion andManagement vol 173 pp 123ndash142 2018

[43] G-Q Zeng J Chen Y-X Dai L-M Li C-W Zheng andM-R Chen ldquoAn improved Design of fractional order PIDcontroller for automatic regulator voltage system based onmulti-objective extremal optimizationrdquo Neurocomputingvol 160 pp 173ndash184 2015

[44] X Xu R Gu F Dai L Qi and S Wan ldquoMulti-objectivecomputation offloading for internet of vehicles in cloud-edgecomputingrdquo Wireless Networks pp 1ndash19 2019

[45] S Wan X Li Y Xue W Lin and X Xu ldquoEfficient com-putation offloading for internet of vehicles in edge comput-ing-assisted 5G networksrdquo e Journal of Supercomputingpp 1ndash30 2019

[46] B Wang S Wan X Zhang and K-K R Choo ldquoA novelindex for assessing the robustness of integrated electricalnetwork and a natural gas networkrdquo IEEE Access vol 6pp 40400ndash40410 2018

[47] Y Yin F Yu Y Xu L Yu and J Mu ldquoNetwork location-aware service recommendation with random walk in cyber-physical systemsrdquo Sensors vol 17 no 9 p 2059 2017

Complexity 13

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Page 11: Parallel LSTM-Based Regional Integrated Energy System ...downloads.hindawi.com/journals/complexity/2019/7414318.pdf · Characteristics of Regional Integrated Energy System “Information

one minute and can meet the requirements of short-termforecasting Further analysis shows that the parallel trainingtime and prediction time of parallel LSTM is greatly im-proved compared to single LSTM due to the combination offive subtasks into one multitask learning In summary al-though the LSTM algorithm is time consuming in trainingtime for short-term prediction the training process doesnot require online completion and the prediction time ofthe LSTM algorithm especially parallel LSTM can fullymeet the requirements of short-term prediction Consid-ering the prediction accuracy the sacrifice of the LSTMalgorithm in computing time is worthwhile

7 Conclusion

In this paper the horizontal and vertical information interactioncharacteristics between multiple energy loads in RIES areconsidered Using LSTM network deep learning algorithm a

parallel LSTM-based multienergy-load information interactiveshort-term energy prediction method is proposed+rough thesimulation of the test data the following conclusions are drawn

(1) +e ldquoinformation interactionrdquo strategy proposed inthis paper can effectively improve the predictionperformance of source-load energy prediction inregional integrated energy systems

(2) Parallel prediction considering horizontal in-teraction mainly affects the prediction effect on theload side while the information sharing of user il-lumination behavior considering vertical interactionmainly affects the improvement of PV predictioneffect

(3) Compared with the conventional prediction algo-rithm the LSTM deep learning algorithm can learnthe deep hidden information of historical dataimprove the prediction accuracy while meetingcalculation time and prove the effectiveness of thealgorithm in the prediction of time series featuresequence

However the performance of the proposed parallelLSTM algorithm may be further enhanced by some evo-lutionary algorithms or multiobjective optimization algo-rithms such as differential evolution algorithm [40]multiobjective extremal optimization [41] and evolutionarymultiobjective optimization algorithms [42ndash47] Of coursesome significant topics are worth researching on the opti-mization of the proposed parallel LSTM by multiobjectiveoptimization

Data Availability

+e data used to support the findings of this study areavailable from the corresponding author upon request

Conflicts of Interest

+e authors declare no potential conflicts of interest

Acknowledgments

+is work was fully supported by the National Key Researchand Development Program of China (no 2018YFB0904200)

References

[1] A Q Huang M L Crow G T Heydt J P Zheng andS J Dale ldquo+e future renewable electric energy delivery andmanagement (FREEDM) system the energy internetrdquo Pro-ceedings of the IEEE vol 99 no 1 pp 133ndash148 2011

[2] I Colak S Sagiroglu G Fulli M Yesilbudak and C CovrigldquoA survey on the critical issues in smart grid technologiesrdquo

Table 5 Comparison of prediction results of three algorithms

Precision Wind power PV power Electric load Heat load Cold load Mean precisionLSTM 08414 08453 08836 09169 09124 08728BPNN 08216 08314 08695 08912 08889 08542ARMIA 07733 07824 08282 08454 08408 08075

Prediction precision ofwind powerPrediction precision ofPV powerPrediction precision ofelectric load

Prediction precision ofhead loadPrediction precision ofcold loadWeight average predictionprecision

07

075

08

085

09

095

1

Pred

ictio

n pr

ecisi

on

BPNN ARMIALSTMPrediction algorithm

Figure 10 Comparison of prediction effects of three algorithms

Table 6 Comparison of computational efficiency of fouralgorithms

Algorithm Training timeaverage (sec)

Prediction timeaverage (sec)

Parallel LSTM 436374 1002Single LSTM 1218337 3662BPNN 1094911 2253ARMIA 54696 215

Complexity 11

Renewable and Sustainable Energy Reviews vol 54 pp 396ndash405 2016

[3] C Tu X He Z Shuai and F Jiang ldquoBig data issues in smartgrid-a reviewrdquo Renewable and Sustainable Energy Reviewsvol 79 pp 1099ndash1107 2017

[4] Y Zhang R Yang K Zhang H Jiang and J J ZhangldquoConsumption behavior analytics-aided energy forecastingand dispatchrdquo IEEE Intelligent Systems vol 32 no 4pp 59ndash63 2017

[5] J Deng F Wang Y Chen and X Zhao ldquoFrom industry 40 toenergy 50 the concept connotation and system framework ofintelligent energy systemsrdquo Journal of Automation vol 41no 12 pp 2003ndash2016 2015

[6] Q Sun L Shi Y Ni D Si and J Zhu ldquoAn enhanced cas-cading failure model integrating data mining techniquerdquoProtection and Control of Modern Power Systems vol 2 2017

[7] D Zheng A T Eseye J Zhang and H Li ldquoShort-term windpower forecasting using a double-stage hierarchical ANFISapproach for energy management in microgridsrdquo Protectionand Control of Modern Power Systems vol 2 pp 1ndash10 2017

[8] Z Li L Ye Y Zhao X Song J Teng and J Jin ldquoShort-termwind power prediction based on extreme learning machinewith error correctionrdquo Protection and Control of ModernPower Systems vol 1 no 1 pp 1ndash8 2016

[9] H Z Wang Z X Lei X Zhang B Zhou and J C Peng ldquoAreview of deep learning for renewable energy forecastingrdquoEnergy Conversion and Management vol 198 pp 1ndash16 2019

[10] H Z Wang G B Wang G Q Li J C Peng and Y T LiuldquoDeep belief network based deterministic and probabilisticwind speed forecasting approachrdquo Applied Energy vol 182pp 80ndash93 2016

[11] M-R Chen G-Q Zeng K-D Lu and J Weng ldquoA two-layernonlinear combination method for short-term wind speedprediction based on ELM ENN and LSTMrdquo IEEE Internet ofings Journal vol 6 no 4 pp 6997ndash7010 2019

[12] H Liu X Mi and Y Li ldquoSmart multi-step deep learningmodel for wind speed forecasting based on variational modedecomposition singular spectrum analysis LSTM networkand ELMrdquo Energy Conversion and Management vol 159pp 54ndash64 2018

[13] J Chen G-Q Zeng W Zhou W Du and K-D Lu ldquoWindspeed forecasting using nonlinear-learning ensemble of deeplearning time series prediction and extremal optimizationrdquoEnergy Conversion and Management vol 165 pp 681ndash6952018

[14] W Kong Z Y Dong Y Jia D J Hill Y Xu and Y ZhangldquoShort-term residential load forecasting based on LSTM re-current neural networkrdquo IEEE Transactions on Smart Gridvol 10 no 1 2017

[15] H Gao W Huang X Yang Y Duan and Y Yin ldquoTowardservice selection for workflow reconfiguration an interface-based computing solutionrdquo Future Generation ComputerSystems vol 87 pp 298ndash311 2018

[16] L Qi Q He F Chen et al ldquoFinding all you need web APIsrecommendation in web of things through keywords searchrdquoIEEE Transactions on Computational Social Systems vol 6no 5 2019

[17] L Qi R Wang C Hu S Li Q He and X Xu ldquoTime-awaredistributed service recommendation with privacy-preserva-tionrdquo Information Sciences vol 480 pp 354ndash364 2019

[18] Y Zhang and Q Yang ldquoAn overview of multi-task learningrdquoNational Science Review vol 5 no 1 pp 34ndash47 2018

[19] A Argyriou M T Evgeniou and M Pontil ldquoConvex multi-task feature learningrdquo Machine Learning vol 73 no 3pp 243ndash272 2008

[20] T Jebara ldquoMultitask sparsity via maximum entropy dis-criminationrdquo Journal of Machine Learning Research vol 12no 1 pp 75ndash110 2011

[21] L Qi Y Chen Y Yuan S Fu X Zhang and X Xu ldquoA QoS-aware virtual machine scheduling method for energy con-servation in cloud-based cyber-physical systemsrdquo WorldWide Web vol 5 pp 1ndash23 2019

[22] R Zhang P Xie C Wang G Liu and S Wan ldquoClassifyingtransportation mode and speed from trajectory data via deepmulti-scale learningrdquo Computer Networks vol 162 p 1068612019

[23] W Wang D Wang H Jia et al ldquoReview of steady-stateanalysis of typical regional integrated energy system under thebackground of energy internetrdquo Proceedings of the CSEE12pp 3292ndash3306 2016

[24] S Ding S Qu Y Xi and S Wan ldquoStimulus-driven andconcept-driven analysis for image caption generationrdquoNeurocomputing 2019

[25] P Vrba V Marik P Siano et al ldquoA review of agent andservice-oriented concepts applied to intelligent energy sys-temsrdquo IEEE Transactions on Industrial Informatics vol 10no 3 pp 1890ndash1903 2014

[26] W Jia C Kang and Q Chen ldquoAnalysis on demand-sideinteractive response capability for power system dispatch in asmart grid frameworkrdquo Electric Power Systems Researchvol 90 pp 11ndash17 2012

[27] Z C Lipton J Berkowitz and C Elkan ldquoA critical review ofrecurrent neural networks for sequence learningrdquo CoRR2015 httpsarxivorgpdf150600019pdf

[28] T N Sainath O Vinyals A Senior and H Sak ldquoCon-volutional long short-term memory fully connected deepneural networksrdquo in Proceedings of the 2015 IEEE In-ternational Conference on Acoustics Speech and Signal Pro-cessing (ICASSP) pp 4580ndash4584 IEEE Brisbane AustraliaApril 2015

[29] Y Guo Z Cheng L Nie YWang J Ma andM KankanhallildquoAttentive long short-term preference modeling for person-alized product searchrdquo ACM Transactions on InformationSystems (TOIS) vol 37 no 2 p 19 2019

[30] A Graves ldquoLong short-term memoryrdquo Neural Computationvol 9 pp 1735ndash1780 1997

[31] F A Gers J Schmidhuber and F Cummins ldquoLearning toforget continual prediction with LSTMrdquo Neural Computa-tion vol 12 no 10 pp 2451ndash2471 2000

[32] M L Seltzer and J Droppo ldquoMulti-task learning in deepneural networks for improved phoneme recognitionrdquo inProceedings of the 2013 IEEE International Conference onAcoustics Speech and Signal Processing pp 6965ndash6969 IEEEVancouver Canada May 2013

[33] S Wan Y Zhao T Wang Z Gu Q H Abbasi andK-K R Choo ldquoMulti-dimensional data indexing and rangequery processing via Voronoi diagram for internet of thingsrdquoFuture Generation Computer Systems vol 91 pp 382ndash3912019

[34] Z Gao D Y Wang S H Wan H Zhang and Y L WangldquoCognitive-inspired class-statistic matching with triple-con-strain for camera free 3D object retrievalrdquo Future GenerationComputer Systems vol 94 pp 641ndash653 2019

[35] C A Kang A R Brandt and L J Durlofsky ldquoOptimaloperation of an integrated energy system including fossil fuel

12 Complexity

power generation CO2 capture and windrdquo Energy vol 36no 12 pp 6806ndash6820 2011

[36] Q Zhang SWan BWang DW Gao and HMa ldquoAnomalydetection based on randommatrix theory for industrial powersystemsrdquo Journal of Systems Architecture vol 95 pp 67ndash742019

[37] Z Gao H Z Xuan H Zhang S Wan and K K R ChooldquoAdaptive fusion and category-level dictionary learningmodel for multi-view human action recognitionrdquo IEEE In-ternet of ings Journal p 1 2019

[38] WWu K Chen Y Qiao and Z Lu ldquoProbabilistic short-termwind power forecasting based on deep neural networksrdquo inProceedings of the 2016 International Conference on Proba-bilistic Methods Applied to Power Systems (PMAPS) pp 1ndash8Beijing China October 2016

[39] K M Powell A Sriprasad W J Cole and T F EdgarldquoHeating cooling and electrical load forecasting for a large-scale district energy systemrdquo Energy vol 74 pp 877ndash8852014

[40] A Peimankar S J Weddell T Jalal and A C LapthornldquoMulti-objective ensemble forecasting with an application topower transformersrdquo Applied Soft Computing vol 68pp 233ndash248 2018

[41] G-Q Zeng J Chen L-M Li et al ldquoAn improved multi-objective population-based extremal optimization algorithmwith polynomial mutationrdquo Information Sciences vol 330pp 49ndash73 2016

[42] Y-L Hu and L Chen ldquoA nonlinear hybrid wind speedforecasting model using LSTM network hysteretic ELM andDifferential Evolution algorithmrdquo Energy Conversion andManagement vol 173 pp 123ndash142 2018

[43] G-Q Zeng J Chen Y-X Dai L-M Li C-W Zheng andM-R Chen ldquoAn improved Design of fractional order PIDcontroller for automatic regulator voltage system based onmulti-objective extremal optimizationrdquo Neurocomputingvol 160 pp 173ndash184 2015

[44] X Xu R Gu F Dai L Qi and S Wan ldquoMulti-objectivecomputation offloading for internet of vehicles in cloud-edgecomputingrdquo Wireless Networks pp 1ndash19 2019

[45] S Wan X Li Y Xue W Lin and X Xu ldquoEfficient com-putation offloading for internet of vehicles in edge comput-ing-assisted 5G networksrdquo e Journal of Supercomputingpp 1ndash30 2019

[46] B Wang S Wan X Zhang and K-K R Choo ldquoA novelindex for assessing the robustness of integrated electricalnetwork and a natural gas networkrdquo IEEE Access vol 6pp 40400ndash40410 2018

[47] Y Yin F Yu Y Xu L Yu and J Mu ldquoNetwork location-aware service recommendation with random walk in cyber-physical systemsrdquo Sensors vol 17 no 9 p 2059 2017

Complexity 13

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Page 12: Parallel LSTM-Based Regional Integrated Energy System ...downloads.hindawi.com/journals/complexity/2019/7414318.pdf · Characteristics of Regional Integrated Energy System “Information

Renewable and Sustainable Energy Reviews vol 54 pp 396ndash405 2016

[3] C Tu X He Z Shuai and F Jiang ldquoBig data issues in smartgrid-a reviewrdquo Renewable and Sustainable Energy Reviewsvol 79 pp 1099ndash1107 2017

[4] Y Zhang R Yang K Zhang H Jiang and J J ZhangldquoConsumption behavior analytics-aided energy forecastingand dispatchrdquo IEEE Intelligent Systems vol 32 no 4pp 59ndash63 2017

[5] J Deng F Wang Y Chen and X Zhao ldquoFrom industry 40 toenergy 50 the concept connotation and system framework ofintelligent energy systemsrdquo Journal of Automation vol 41no 12 pp 2003ndash2016 2015

[6] Q Sun L Shi Y Ni D Si and J Zhu ldquoAn enhanced cas-cading failure model integrating data mining techniquerdquoProtection and Control of Modern Power Systems vol 2 2017

[7] D Zheng A T Eseye J Zhang and H Li ldquoShort-term windpower forecasting using a double-stage hierarchical ANFISapproach for energy management in microgridsrdquo Protectionand Control of Modern Power Systems vol 2 pp 1ndash10 2017

[8] Z Li L Ye Y Zhao X Song J Teng and J Jin ldquoShort-termwind power prediction based on extreme learning machinewith error correctionrdquo Protection and Control of ModernPower Systems vol 1 no 1 pp 1ndash8 2016

[9] H Z Wang Z X Lei X Zhang B Zhou and J C Peng ldquoAreview of deep learning for renewable energy forecastingrdquoEnergy Conversion and Management vol 198 pp 1ndash16 2019

[10] H Z Wang G B Wang G Q Li J C Peng and Y T LiuldquoDeep belief network based deterministic and probabilisticwind speed forecasting approachrdquo Applied Energy vol 182pp 80ndash93 2016

[11] M-R Chen G-Q Zeng K-D Lu and J Weng ldquoA two-layernonlinear combination method for short-term wind speedprediction based on ELM ENN and LSTMrdquo IEEE Internet ofings Journal vol 6 no 4 pp 6997ndash7010 2019

[12] H Liu X Mi and Y Li ldquoSmart multi-step deep learningmodel for wind speed forecasting based on variational modedecomposition singular spectrum analysis LSTM networkand ELMrdquo Energy Conversion and Management vol 159pp 54ndash64 2018

[13] J Chen G-Q Zeng W Zhou W Du and K-D Lu ldquoWindspeed forecasting using nonlinear-learning ensemble of deeplearning time series prediction and extremal optimizationrdquoEnergy Conversion and Management vol 165 pp 681ndash6952018

[14] W Kong Z Y Dong Y Jia D J Hill Y Xu and Y ZhangldquoShort-term residential load forecasting based on LSTM re-current neural networkrdquo IEEE Transactions on Smart Gridvol 10 no 1 2017

[15] H Gao W Huang X Yang Y Duan and Y Yin ldquoTowardservice selection for workflow reconfiguration an interface-based computing solutionrdquo Future Generation ComputerSystems vol 87 pp 298ndash311 2018

[16] L Qi Q He F Chen et al ldquoFinding all you need web APIsrecommendation in web of things through keywords searchrdquoIEEE Transactions on Computational Social Systems vol 6no 5 2019

[17] L Qi R Wang C Hu S Li Q He and X Xu ldquoTime-awaredistributed service recommendation with privacy-preserva-tionrdquo Information Sciences vol 480 pp 354ndash364 2019

[18] Y Zhang and Q Yang ldquoAn overview of multi-task learningrdquoNational Science Review vol 5 no 1 pp 34ndash47 2018

[19] A Argyriou M T Evgeniou and M Pontil ldquoConvex multi-task feature learningrdquo Machine Learning vol 73 no 3pp 243ndash272 2008

[20] T Jebara ldquoMultitask sparsity via maximum entropy dis-criminationrdquo Journal of Machine Learning Research vol 12no 1 pp 75ndash110 2011

[21] L Qi Y Chen Y Yuan S Fu X Zhang and X Xu ldquoA QoS-aware virtual machine scheduling method for energy con-servation in cloud-based cyber-physical systemsrdquo WorldWide Web vol 5 pp 1ndash23 2019

[22] R Zhang P Xie C Wang G Liu and S Wan ldquoClassifyingtransportation mode and speed from trajectory data via deepmulti-scale learningrdquo Computer Networks vol 162 p 1068612019

[23] W Wang D Wang H Jia et al ldquoReview of steady-stateanalysis of typical regional integrated energy system under thebackground of energy internetrdquo Proceedings of the CSEE12pp 3292ndash3306 2016

[24] S Ding S Qu Y Xi and S Wan ldquoStimulus-driven andconcept-driven analysis for image caption generationrdquoNeurocomputing 2019

[25] P Vrba V Marik P Siano et al ldquoA review of agent andservice-oriented concepts applied to intelligent energy sys-temsrdquo IEEE Transactions on Industrial Informatics vol 10no 3 pp 1890ndash1903 2014

[26] W Jia C Kang and Q Chen ldquoAnalysis on demand-sideinteractive response capability for power system dispatch in asmart grid frameworkrdquo Electric Power Systems Researchvol 90 pp 11ndash17 2012

[27] Z C Lipton J Berkowitz and C Elkan ldquoA critical review ofrecurrent neural networks for sequence learningrdquo CoRR2015 httpsarxivorgpdf150600019pdf

[28] T N Sainath O Vinyals A Senior and H Sak ldquoCon-volutional long short-term memory fully connected deepneural networksrdquo in Proceedings of the 2015 IEEE In-ternational Conference on Acoustics Speech and Signal Pro-cessing (ICASSP) pp 4580ndash4584 IEEE Brisbane AustraliaApril 2015

[29] Y Guo Z Cheng L Nie YWang J Ma andM KankanhallildquoAttentive long short-term preference modeling for person-alized product searchrdquo ACM Transactions on InformationSystems (TOIS) vol 37 no 2 p 19 2019

[30] A Graves ldquoLong short-term memoryrdquo Neural Computationvol 9 pp 1735ndash1780 1997

[31] F A Gers J Schmidhuber and F Cummins ldquoLearning toforget continual prediction with LSTMrdquo Neural Computa-tion vol 12 no 10 pp 2451ndash2471 2000

[32] M L Seltzer and J Droppo ldquoMulti-task learning in deepneural networks for improved phoneme recognitionrdquo inProceedings of the 2013 IEEE International Conference onAcoustics Speech and Signal Processing pp 6965ndash6969 IEEEVancouver Canada May 2013

[33] S Wan Y Zhao T Wang Z Gu Q H Abbasi andK-K R Choo ldquoMulti-dimensional data indexing and rangequery processing via Voronoi diagram for internet of thingsrdquoFuture Generation Computer Systems vol 91 pp 382ndash3912019

[34] Z Gao D Y Wang S H Wan H Zhang and Y L WangldquoCognitive-inspired class-statistic matching with triple-con-strain for camera free 3D object retrievalrdquo Future GenerationComputer Systems vol 94 pp 641ndash653 2019

[35] C A Kang A R Brandt and L J Durlofsky ldquoOptimaloperation of an integrated energy system including fossil fuel

12 Complexity

power generation CO2 capture and windrdquo Energy vol 36no 12 pp 6806ndash6820 2011

[36] Q Zhang SWan BWang DW Gao and HMa ldquoAnomalydetection based on randommatrix theory for industrial powersystemsrdquo Journal of Systems Architecture vol 95 pp 67ndash742019

[37] Z Gao H Z Xuan H Zhang S Wan and K K R ChooldquoAdaptive fusion and category-level dictionary learningmodel for multi-view human action recognitionrdquo IEEE In-ternet of ings Journal p 1 2019

[38] WWu K Chen Y Qiao and Z Lu ldquoProbabilistic short-termwind power forecasting based on deep neural networksrdquo inProceedings of the 2016 International Conference on Proba-bilistic Methods Applied to Power Systems (PMAPS) pp 1ndash8Beijing China October 2016

[39] K M Powell A Sriprasad W J Cole and T F EdgarldquoHeating cooling and electrical load forecasting for a large-scale district energy systemrdquo Energy vol 74 pp 877ndash8852014

[40] A Peimankar S J Weddell T Jalal and A C LapthornldquoMulti-objective ensemble forecasting with an application topower transformersrdquo Applied Soft Computing vol 68pp 233ndash248 2018

[41] G-Q Zeng J Chen L-M Li et al ldquoAn improved multi-objective population-based extremal optimization algorithmwith polynomial mutationrdquo Information Sciences vol 330pp 49ndash73 2016

[42] Y-L Hu and L Chen ldquoA nonlinear hybrid wind speedforecasting model using LSTM network hysteretic ELM andDifferential Evolution algorithmrdquo Energy Conversion andManagement vol 173 pp 123ndash142 2018

[43] G-Q Zeng J Chen Y-X Dai L-M Li C-W Zheng andM-R Chen ldquoAn improved Design of fractional order PIDcontroller for automatic regulator voltage system based onmulti-objective extremal optimizationrdquo Neurocomputingvol 160 pp 173ndash184 2015

[44] X Xu R Gu F Dai L Qi and S Wan ldquoMulti-objectivecomputation offloading for internet of vehicles in cloud-edgecomputingrdquo Wireless Networks pp 1ndash19 2019

[45] S Wan X Li Y Xue W Lin and X Xu ldquoEfficient com-putation offloading for internet of vehicles in edge comput-ing-assisted 5G networksrdquo e Journal of Supercomputingpp 1ndash30 2019

[46] B Wang S Wan X Zhang and K-K R Choo ldquoA novelindex for assessing the robustness of integrated electricalnetwork and a natural gas networkrdquo IEEE Access vol 6pp 40400ndash40410 2018

[47] Y Yin F Yu Y Xu L Yu and J Mu ldquoNetwork location-aware service recommendation with random walk in cyber-physical systemsrdquo Sensors vol 17 no 9 p 2059 2017

Complexity 13

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Page 13: Parallel LSTM-Based Regional Integrated Energy System ...downloads.hindawi.com/journals/complexity/2019/7414318.pdf · Characteristics of Regional Integrated Energy System “Information

power generation CO2 capture and windrdquo Energy vol 36no 12 pp 6806ndash6820 2011

[36] Q Zhang SWan BWang DW Gao and HMa ldquoAnomalydetection based on randommatrix theory for industrial powersystemsrdquo Journal of Systems Architecture vol 95 pp 67ndash742019

[37] Z Gao H Z Xuan H Zhang S Wan and K K R ChooldquoAdaptive fusion and category-level dictionary learningmodel for multi-view human action recognitionrdquo IEEE In-ternet of ings Journal p 1 2019

[38] WWu K Chen Y Qiao and Z Lu ldquoProbabilistic short-termwind power forecasting based on deep neural networksrdquo inProceedings of the 2016 International Conference on Proba-bilistic Methods Applied to Power Systems (PMAPS) pp 1ndash8Beijing China October 2016

[39] K M Powell A Sriprasad W J Cole and T F EdgarldquoHeating cooling and electrical load forecasting for a large-scale district energy systemrdquo Energy vol 74 pp 877ndash8852014

[40] A Peimankar S J Weddell T Jalal and A C LapthornldquoMulti-objective ensemble forecasting with an application topower transformersrdquo Applied Soft Computing vol 68pp 233ndash248 2018

[41] G-Q Zeng J Chen L-M Li et al ldquoAn improved multi-objective population-based extremal optimization algorithmwith polynomial mutationrdquo Information Sciences vol 330pp 49ndash73 2016

[42] Y-L Hu and L Chen ldquoA nonlinear hybrid wind speedforecasting model using LSTM network hysteretic ELM andDifferential Evolution algorithmrdquo Energy Conversion andManagement vol 173 pp 123ndash142 2018

[43] G-Q Zeng J Chen Y-X Dai L-M Li C-W Zheng andM-R Chen ldquoAn improved Design of fractional order PIDcontroller for automatic regulator voltage system based onmulti-objective extremal optimizationrdquo Neurocomputingvol 160 pp 173ndash184 2015

[44] X Xu R Gu F Dai L Qi and S Wan ldquoMulti-objectivecomputation offloading for internet of vehicles in cloud-edgecomputingrdquo Wireless Networks pp 1ndash19 2019

[45] S Wan X Li Y Xue W Lin and X Xu ldquoEfficient com-putation offloading for internet of vehicles in edge comput-ing-assisted 5G networksrdquo e Journal of Supercomputingpp 1ndash30 2019

[46] B Wang S Wan X Zhang and K-K R Choo ldquoA novelindex for assessing the robustness of integrated electricalnetwork and a natural gas networkrdquo IEEE Access vol 6pp 40400ndash40410 2018

[47] Y Yin F Yu Y Xu L Yu and J Mu ldquoNetwork location-aware service recommendation with random walk in cyber-physical systemsrdquo Sensors vol 17 no 9 p 2059 2017

Complexity 13

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Page 14: Parallel LSTM-Based Regional Integrated Energy System ...downloads.hindawi.com/journals/complexity/2019/7414318.pdf · Characteristics of Regional Integrated Energy System “Information

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom