performance*measuresfor*...

1
This work is supported by the US Department of Energy, the Los Angeles Department of Water and Power, and the NaAonal Science FoundaAon. The views of authors expressed herein do not necessarily reflect those of the sponsors. Performance Measures for Electricity Consumption Prediction Saima Aman (Advisors: Yogesh Simmhan and Viktor K. Prasanna) http://ganges.usc.edu riskadjusted improvement over baseline factor in vola6lity of model with respect to baseline PredicAon Bias Scale Independence Reliability Cost VolaAlity understand the frequency of over or underpredic6on underpredic6on might miss the peak compare across different scales (unlike MAE, RMSE) address diversity in customers how oEen the model performs beFer than a baseline or within an error threshold quan6fy the cost of collec6ng data, training and applying a model for predic6on Domain Bias Percentage Error (DBPE) An asymmetric loss funcAon is used to assign different costs to over and under predicAons. These costs are applicaAonspecific. (Reduces to MAPE when costs are same) Problem: EvaluaAon of KWh predicAon Coefficient of Varia:on of RMSE (CVRMSE) The root mean square error is divided by the mean of observed values. The normalized RMSE can then be used to compare across scales. Reliability, REL Measures the count of performances less than the error threshold. Rela:ve Improvement, RIM Measures the count of performances beVer than the baseline. Need for novel Performance Measures Data Cost, DC The number of unique values of all features in the model. Compute Cost, CC The Ame in seconds required to train a model Normalized Model cost, C = f(DC, CC)/m CostBenefit Metric, CBM Measures the relaAve benefit of using a model with respect to normalized cost. Vola:lity Adjusted Benefit BVM. It is adapted from the Sharpe raAo used in the finance domain for measuring benefit to risk raAo. 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 Annual Mean DoW Mean DoY Mean TS 1 wk ahead TS 2 wk ahead Ts 3 wk ahead TS 4wk ahead RT CVRMSE DBPE (0.5, 1.5) Ini:al Results 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 TS 1 wk ahead TS 2 wk ahead Ts 3 wk ahead TS 4wk ahead RT RIM REL BVM Mo:va:on: Dynamic Demand Response (D 2 R) Dynamic decision making for start Ame duraAon depth (kWh) customer selecAon curtailment strategy selecAon Error Measures (smaller is beVer) Goodness Measures (larger is beVer)

Upload: others

Post on 28-May-2020

2 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Performance*Measuresfor* Electricity*Consumption*Predictionsaimacs.github.io/pubs/2013-csrr-poster.pdf · This%work%is%supported%by%the%US%Departmentof%Energy,%the%Los%Angeles%Departmentof%Water%and%Power,%and%the%Naonal%

This  work  is  supported  by  the  US  Department  of  Energy,  the  Los  Angeles  Department  of  Water  and  Power,  and  the  NaAonal  Science  FoundaAon.  The  views  of  authors  expressed  herein  do  not  necessarily  reflect  those  of  the  sponsors.  

Performance  Measures  for  Electricity  Consumption  Prediction  

Saima Aman (Advisors: Yogesh Simmhan and Viktor K. Prasanna)

http://ganges.usc.edu  

•  risk-­‐adjusted  improvement  over  baseline  •  factor  in  vola6lity  of  model  with  respect  to  baseline  

PredicAon  Bias  

Scale  Independence  

Reliability  

Cost  

VolaAlity  

•  understand  the  frequency  of  over-­‐  or  under-­‐predic6on  •  under-­‐predic6on  might  miss  the  peak  

•  compare  across  different  scales  (unlike  MAE,  RMSE)  •  address  diversity  in  customers  

•  how  oEen  the  model  performs  beFer  than  a  baseline  or  within  an  error  threshold  

•  quan6fy  the  cost  of  collec6ng  data,  training  and  applying  a  model  for  predic6on  

 

Domain  Bias  Percentage  Error  (DBPE)  An  asymmetric  loss  funcAon  is  used  to  assign  different  costs  to  over  and  under  predicAons.  These  costs  are  applicaAon-­‐specific.  (Reduces  to  MAPE  when  costs  are  same)      

Problem:  EvaluaAon  of  KWh  predicAon  

Coefficient  of  Varia:on  of  RMSE  (CV-­‐RMSE)  The  root  mean  square  error  is  divided  by  the  mean  of  observed  values.  The  normalized  RMSE  can  then  be  used  to  compare  across  scales.      

Reliability,  REL    Measures  the  count  of  performances  less  than  the  error  threshold.  

Rela:ve  Improvement,  RIM    Measures  the  count  of  performances  beVer  than  the  baseline.  

Need  for  novel  Performance  Measures  

Data  Cost,  DC    The  number  of  unique  values  of  all  features  in  the  model.  Compute  Cost,  CC  The  Ame  in  seconds  required  to  train  a  model    Normalized  Model  cost,  C  =  f(DC,  CC)/m

Cost-­‐Benefit  Metric,  CBM  Measures  the  relaAve  benefit  of  using  a  model  with  respect  to  normalized  cost.  

Vola:lity  Adjusted  Benefit-­‐  BVM.  It  is  adapted  from  the  Sharpe  raAo  used  in  the  finance  domain  for  measuring  benefit  to    risk  raAo.  

0  

0.02  

0.04  

0.06  

0.08  

0.1  

0.12  

0.14  

Annual  Mean   DoW  Mean   DoY  Mean   TS  1  wk  ahead  TS  2  wk  ahead  Ts  3  wk  ahead  TS  4-­‐wk  ahead   RT  

CVRMSE   DBPE  (0.5,  1.5)  

Ini:al  Results  

0.0  0.1  0.2  0.3  0.4  0.5  0.6  0.7  0.8  0.9  1.0  

TS  1  wk  ahead  

TS  2  wk  ahead  

Ts  3  wk  ahead   TS  4-­‐wk  ahead  

RT  

RIM  

REL  

BVM  

Mo:va:on:  Dynamic  Demand  Response  (D2R)  

Dynamic  decision  making  for    •  start  Ame  •  duraAon  •  depth  (kWh)  •  customer  selecAon  •  curtailment  strategy  selecAon  

Error  Measures  (smaller  is  beVer)   Goodness  Measures  (larger  is  beVer)