PROCESS STATE IDENTIFICATION AND MODELING IN AFLUIDIZED BED ENERGY PLANT BY USING ARTIFICIAL

NEURAL NETWORKS

MIKA LIUKKONEN1,*, EERO HÄLIKKÄ2, REIJO KUIVALAINEN2, YRJÖ HILTUNEN1

1Department of Environmental Science, University of Kuopio, Finland2Foster Wheeler Power Group, Varkaus, Finland

*E-mail: mika.liukkonen@uku.fi, Tel.: +358 40 351 0644

ABSTRACT

The aims to reduce harmful emissions are influencing increasingly the modern-dayproduction of energy, while higher objectives are set also for the efficiency of combus-tion processes. Therefore it is necessary to develop such data analysis and modelingmethods that can respond to these demands. This paper presents an overview of how theformation of nitrogen oxides (NOx) in a circulating fluidized bed (CFB) boiler wasmodeled by using a sub-model -based artificial neural network (ANN) approach. In thisapproach, the process data is processed first by using a self-organizing map (SOM) andk-means clustering to generate subsets representing the separate process states in theboiler. These primary process states represent the higher level process conditions in thecombustion, and can include for example start-ups, shutdowns, and idle times in addi-tion to the normal process flow. However, the primary states of process may containsecondary states that represent more subtle phenomena in the process, which are moredifficult to observe. The data from secondary combustion conditions can involve infor-mation on e.g. instabilities in the process. In this study, the aims were to identify thesesecondary process states and to show that in some cases the simulation accuracy can beimproved by creating secondary sub-models. The results show that the approach pre-sented can be a fruitful way to get new information from combustion processes.

Keywords: Modeling, Fluidized bed, Neural networks, Self-organizing map, Multilayerperceptron, Simulation, Process state

1. INTRODUCTION

Nowadays the efficiency of energy plants is considered an important topic due to envi-ronmental issues and increasing production costs. Efficient combustion of fuels withlower emissions is a challenging task in the modern-day production of energy. Fortu-nately, process data can involve important information on the behavior of the processand on different phenomena that affect the emissions and the energy efficiency of acombustion process. This information can be valuable when optimizing the process. Forthis reason, new methods are needed for data processing, analysis and modeling.

Artificial neural networks (ANN) have shown their usability and power in the modelingof industrial processes (Harding et al., 2006; Haykin, 1999; Meireles et al., 2003; Muj-taba & Hussain, 2001). ANNs have provided serviceable applications in diverse fieldsof industry, for example in energy production, chemical and electronics industry, andeven in waste water treatment (Heikkinen et al., 2008a; Heikkinen et al., 2008b; Heik-kinen et al., 2005; Liukkonen et al., 2008a; Liukkonen et al., 2008b). The variety ofneural network applications is wide due to their strong advantages including flexibility,nonlinearity, adaptivity, applicability, a high computing power and a high tolerance offaults (Haykin, 1999; Kohonen, 2001; Reed & Marks II, 1999). These benefits makeANNs a valuable choice for modeling method in industrial processes.

We have showed earlier (Heikkinen et al., 2005) that different states of the fluidized bedcombustion process can be discovered in process data. In reality these states can be forinstance start-ups, shut-downs, idle times and different states of normal process flow.The behavior of the process, e.g. the quantities of different emission components, can bediverse between these conditions. However, these upper level process states also includesecondary process states, where for example the bed temperature is unstable or thesteam flow is lower than usual. It is momentous to learn to identify these states as well,because the performance of the process can fluctuate also in a smaller scale but regard-less in a way that has an effect on the model accuracy.

Studying sub-models in parallel to more generic models opens an interesting view tomodern-day process analysis. This is because process states and their correspondingsub-models can include valuable information on the performance of the process, as ourearlier results concerning the wave soldering and the activated sludge treatment processindicate (Heikkinen et al., 2008a; Liukkonen et al., 2008b). The sub-model -based ap-proach is a realistic option for instance in cases where it seems apparent that less detect-able but still important phenomena are hidden under the generic behavior of the data. Inspite of being more difficult to recognize, these phenomena can have a substantial effecton certain events in the combustion process. Fortunately, this kind of underlying infor-mation can be exposed by identifying different process states and creating sub-models,progressing from universal to more detailed models.

In this study, ANNs were used to identify the different states of a circulating fluidizedbed (CFB) process and to create sub-models where the nitrogen oxide content of theflue gas was simulated. The obtained sub-models were then compared to the generic

process model to see whether the accuracy of simulation could be improved by usingthis method. In the approach used, self-organizing maps (Kohonen, 2001), k-meansclustering (MacQueen, 1967) and multilayer perceptrons (Haykin, 1999) were com-bined in a sequential manner to form a hybrid ANN method that benefits the good cha-racteristics of all these methods.

2. PROCESS AND DATA

Fluidized bed combustion is a widely-used combustion technology used in power plantsand designed primarily for solid fuels. A typical circulating fluidized bed (CFB) boilerconsists of a combustion chamber, a separator and a return leg for the recirculation ofthe bed particles. The fluidized bed is characteristically composed of sand, fuel ash anda matter for capturing the sulfur. This mixture is fluidized by the primary combustion airbrought in from the bottom of the chamber. Because of high fluidizing velocities, thebed particles are persistently moving with the flue gases. The particles are driventhrough the main combustion chamber into a separator, where the larger particles areextracted and returned to the combustion chamber. Meanwhile, the finer particles areseparated from the circulation and removed from the flue gases by a bag-house filter oran electrostatic precipitator located downstream from the boiler’s convection section.

One of the advantages in fluidized bed combustion is the large heat capacity of the bed,which ensures steady combustion. Only the start-ups involve the use of supporting fuelssuch as oil or gas. The purpose of the strong turbulence in the circulating fluidized bedis to support the mixing and combustion of fuel. The typical combustion temperature inCFB boilers is between 850 and 900 °C. The process data were compressed for theanalysis by using a moving average to comprise 10 000 data rows with a five minutetime interval, the number of variables being 36. In addition to the averaged variables,also the standard deviations were exploited in the analysis to indicate the possible insta-bilities in the process.

3. METHODS

3.1 Self-organizing maps (SOM)

Kohonen’s self-organizing map (SOM) (Kohonen, 2001) is a well-known unsupervisedartificial neural network algorithm. The common purpose of SOM is to facilitate dataanalysis by mapping n-dimensional input vectors to the neurons for example in a two-dimensional lattice (map). The map consists of classes, or neurons, where the data rowsare mapped to get a generalization of the original data. The map reflects variations inthe statistics of the data set and selects common features which approximate to thedistribution of the data samples. On the SOM, the input vectors with common featuresare associated with the same or neighboring neurons, which preserves the topologicalorganization of the input data. The common properties of a map neuron can be pre-sented with an n-dimensional, neuron-specific reference vector (prototype vector). Thesize of the map, or the number of neurons, can be altered depending on the purpose; themore neurons, the more details appear.

The SOM analysis is premised on unsupervised learning. At first, the preliminary refer-ence vectors are initialized randomly by sampling their values from an even distributionwhose limits are defined by the input data. During learning the input vectors are thengrouped one by one into best matching units (BMU) on the map. The BMU is the neu-ron whose reference vector has the smallest n-dimensional Euclidean distance to theinput vector. At the same time, the nearest neighbors of the activated neuron becomelikewise activated according to a predefined neighborhood function (e.g. Gaussiandistribution) that is dependent on the network topology. At the final phase, the referencevectors of all activated neurons are updated.

In this study, the SOM was used as a pre-processor to compress information, to removenoise, and to visualize the data. The fluidized bed boiler data were coded into inputs fora self-organizing network, and a SOM having 384 neurons in a 24 x 16 hexagonal ar-rangement was constructed. The linear initialization and batch training algorithm wereused in training, and the neighborhood function was Gaussian. The map was taught with10 epochs, and the initial neighborhood had the value of 6. The SOM Toolbox(http://www.cis.hut.fi/projects/somtoolbox/) was used in the analysis under a Matlabversion 7.6 software (Mathworks Inc., Natick, MA, USA, 2008) platform.

3.2 Identification of process states

K-means clustering (MacQueen, 1967) was used to cluster the SOM reference vectors.The algorithm is initialized by randomly defining k cluster centers, and then directingeach data point to the cluster whose mean value is closest to it. The Euclidean distanceis generally used as a distance measure in these comparisons. At the next stage, themean vectors of the data points assimilated to each cluster are calculated and used asnew centers in an iterative process. Clustering was performed twice, first to reveal theprimary process states, and secondly to discover the secondary process states within theprimary states.

There are several computational methods to determine the optimal number of clusters,one of the mostly used being the Davies-Bouldin -index (Davies & Bouldin, 1979).Small values of the DB-index refer to compact clusters whose centers are far from eachother. Hence the optimal number of clusters is the number where the index reaches itsminimum. The computation of the optimal cluster structure is useful because thus it isnot necessary to know the clusters beforehand.

The difference between two separate cluster center vectors represents the factors thatseparate the clusters. Therefore, this operation can be used to identify the reasons for thedetermination of different process states. In practice, this calculation is performed sothat the comparable individual vector components of the center vectors are subtractedfrom each other. This produces a vector of differences, which can be used to identify theclusters as process states.

3.3 Multilayer perceptrons (MLP)

Multilayer perceptrons (MLP) are a widely-used (Freeman & Skapura, 1991; Haykin,1999; Hecht-Nielsen, 1990; Meireles et al., 2003), supervised ANN methodology thatcan be used for example to create prediction models. MLPs consist of processing ele-ments and connections. The processing elements are comprised of at least three layers:an input layer, one or more hidden layers, and an output layer. In training, the inputs areprocessed forward through consecutive layers of neurons. At first, the input layer distri-butes the input signals to the first hidden layer. Next, the neurons in the hidden layerprocess the inputs based on given weights, which can either weaken or strengthen theeffect of each input. The weights are determined by a supervised learning process,which means learning from examples, or data samples. The inputs are next processed bya transfer function, which is usually nonlinear, providing the method with the ability todistinguish data that are not linearly separable. After that, the hidden neurons transferthe outcome as a linear combination to the next layer, which is generally the outputlayer. At the last stage, the performance of the model is evaluated with an independentvalidation data set. This is done by simulating the known data samples by using themodel, and comparing the simulated output values with the real ones by using somemodel performance measure.

MLP neural networks must be trained to a given problem. One of the mostly used train-ing techniques is the back-propagation (Werbos, 1994) algorithm. In back-propagation,the network’s output value is compared with the original sample to compute the value ofa predefined error function. In the iterative training process the network weights areadjusted to a level that minimizes the error value between the actual and expected out-put for all input patterns.

A MLP network consisting of an input layer, an output layer and a hidden layer with 15hidden neurons was used to simulate the flue gas nitrogen oxide (NOx) content in theprimary and secondary process states defined earlier by clustering. Variance scaling wasused for pre-processing the data. The pre-processed data was divided into training,training test and validation sets. The training data set, being 60 % of the total 10 000samples was used for training the network, while 20 % of the data set was put to theseparate test set to be used in the back-propagation error calculations. The validationdata set included the remaining 20 % of the samples, and was used as an independenttest set for testing the model.

The parameters for the neural network were determined experimentally. The hyperbolictangent sigmoid (tansig) transfer function was used in the hidden layer, and the linear(purelin) transfer function in the output layer. The resilient back-propagation (trainrp)algorithm was operated in the training procedure, and the mean squared error (mse) asthe error function in training. Matlab version 7.6 (Mathworks Inc., Natick, MA, USA)software with the Neural Network Toolbox (version 6.0) was used in data processing. Atthe final stage, the prediction performances of the primary and secondary models werecompared to prove the usefulness of the presented method. Index of agreement (Will-mott, 1981) was used as the model performance measure.

4. RESULTS AND DISCUSSION

In the analysis, the behavior of the data showed strong distribution of the process eventsinto three principal process states. These states are defined best by high, low and me-dium steam flow, as can be seen in Figure 1. This kind of principal clustering behaviorcan be referred to as separating into primary process states. However, the strong group-ing of the data samples caused by the characteristics of the data means that some inter-esting phenomena can remain undetected behind those strongly defined states. Despitetheir smaller visibility, these phenomena can be important considering the combustionprocess and certain events in it. Fortunately, this sort of underlying information can berevealed by identifying secondary process states as presented in Figure 1. The identifi-cation is possible by comparing the cluster center vectors as shown in the figure.

Figure 1. Example on the identification of primary and secondary process states on the SOM. The bargraphs represent the difference between the cluster center vector of the secondary process state and thecluster center vector of the primary process state (high steam flow). A considerable positive difference ofstandard deviations indicates instability in the corresponding variable.

The Figure 2 and Table 1 present the NOx prediction performance of the main modeland the sub-models of the levels 1 and 2. The accuracy of the NOx simulations im-proved by creating secondary states of process and their sub-models, as can be seen inFigure 2 and Table 1. This means that in some cases the model can be improved gradu-ally by getting deeper into the process states and their corresponding sub-models. In thisrespect, the results support our previous findings (Heikkinen et al., 2008a; Liukkonen etal., 2008b) concerning the wave soldering and the activated sludge treatment processes.

The results suggest that sometimes it is reasonable to perform clustering in severalseparate steps to reveal the underlying phenomena that may otherwise not be recog-nized. These hidden phenomena are not automatically perceivable without creating sub-models, but may still have a significant effect on certain factors in the process. In otherwords, the model of a more general level can be sufficient in certain situations, but amore detailed model can be more accurate in describing other phenomena.

Figure 2. Simulated vs. observed flue gas NOx content in different models. a) indicates the generic mainmodel where the whole data is involved, b) is the level 1 sub-model (Cluster III, high steam flow), and c)is the level 2 sub-model (Cluster 7). Solid line describes least squares and dotted line perfect fitting.

a) Main model b) Level 1 Sub-model (Cluster III)

c) Level 2 Sub-model (Cluster 7)

0.950 0.956 0.979

Table 1. The goodness (index of agreement) of models a, b and c presented in Figure 2.

In summary, the data analysis scheme used is illustrated in Figure 3. Firstly, the raw dataare pre-processed. Preprocessing includes all the necessary actions, such as data filter-ing and normalization of variables, to prepare the data for modeling. Next, the SOM andk-means clustering are used and combined with expert process knowledge to obtain theprimary process states and their MLP sub-models. After this, the secondary states ofprocess and their corresponding sub-models are formed using the same approach.

Figure 3. The data processing chain in a nutshell.

The results show that the universal-to-detailed data analysis method presented can beexcellent in cases where high simulation accuracies are required. In addition, the gra-dual penetrating analysis ensures that the best model is always found for each case inspite of its specific level. The classification of data samples into different subcategories,or process states, provides extra accuracy to the model. Furthermore, the ability of themethod to reveal nonlinear and multivariate interactions entails additional value to themodel. For these reasons, the data analysis method presented offers a powerful option tomodel industrial processes.

5. CONCLUSIONS

Developing new methods for the modeling of combustion processes is important be-cause there is a growing need for optimizing the production of energy due to increasingenvironmental issues and higher demands for process efficiency. The results representedin this paper show that the modeling approach used is a fruitful way to model the circu-lating fluidized bed process and to simulate the process emissions.

REFERENCES

DAVIES, D.L., BOULDIN, D.W. (1979). A Cluster Separation Measure. IEEE Trans. Pattern Recogni-tion and Machine Intelligence. Vol. 1, No. 2, pp. 224–227.

FREEMAN, J., SKAPURA, D. (1991). Neural Networks Algorithms, Application, and ProgrammingTechniques. Addison-Wesley Publishing Company, Menlo Park, California.

HARDING, J.A., SHAHBAZ, M., SRINIVAS, KUSIAK, A. (2006). Data Mining in Manufacturing: AReview. Journal of Manufacturing Science and Engineering. Vol. 128, No. 4, pp. 969–976.

HAYKIN, S. (1999). Neural Networks: A Comprehensive Foundation. Prentice Hall, Upper Saddle River,NJ.

HECHT-NIELSEN, R. (1990). Neurocomputing. Addison-Wesley Publishing, San Diego, California.

HEIKKINEN, M., HEIKKINEN, T., HILTUNEN, Y. (2008a). Process States and Their Submodels UsingSelf-Organizing Maps in an Activated Sludge Treatment Plant. In: Proc. 48th Scandinavian Conferenceon Simulation and Modeling (SIMS 2008) [CD-ROM], Oslo University College.

HEIKKINEN, M., NURMINEN, V., HILTUNEN, T., HILTUNEN, Y. (2008b). A Modeling and Optimiza-tion Tool for the Expandable Polystyrene Batch Process. Chemical Product and Process Modeling. Vol. 3,No. 1, Article 3.

HEIKKINEN, M., KETTUNEN, A., NIEMITALO, E., KUIVALAINEN, R., HILTUNEN, Y. (2005).SOM-based method for process state monitoring and optimization in fluidized bed energy plant. In: Duch,W., Kacprzyk, J., Oja, E., Zadro ny, S. (eds.), Lecture Notes in Computer Science 3696, Springer-VerlagBerlin Heidelberg, pp. 409-414.

KOHONEN, T. (2001). Self-Organizing Maps. Springer-Verlag, Berlin Heidelberg New York, 3rd ed.

LIUKKONEN, M., HILTUNEN, T., HAVIA, E., LEINONEN, H., HILTUNEN, Y. (2008a). Modeling ofSoldering Quality by Using Artificial Neural Networks. IEEE Trans. Electronics Packaging Manufactur-ing, accepted for publication.

LIUKKONEN, M., HAVIA, E., LEINONEN, H., HILTUNEN, Y. (2008b). Application of Self-Organizing Maps in Analysis of Wave Soldering Process. Expert Systems with Applications, article inpress.

MACQUEEN, J. (1967). Some methods for classification and analysis of multivariate observations. InProc. Fifth Berkeley Symposium on Mathematical Statistics and Probability, Vol. I: Statistics. Universityof California Press, Berkeley and Los Angeles, pp. 281–297.

MEIRELES, M.R.G., ALMEIDA, P.E.M., SIMÕES, M.G. (2003). A Comprehensive Review for Industri-al Applicability of Artificial Neural Networks. IEEE Trans. Industrial Electronics, Vol. 50, No. 3, pp.585–601.

MUJTABA, I.M., HUSSAIN, M.A., eds. (2001). Application of Neural Network and Other LearningTechnologies in Process Engineering. Imperial College Press London, UK.

REED, R.D., MARKS II, R.J. (1999). Neural Smithing: Supervised Learning in Feedforward ArtificialNeural Networks. MIT Press, Cambridge, Massachusetts.

WERBOS, P.J. (1994). The Roots of Backpropagation: From Ordered Derivatives to Neural Networksand Political Forecasting. John Wiley & Sons, New York.

WILLMOTT, C.J. (1981). On the validation of models. Physical Geography. Vol. 2, pp. 184–194.