Context-Aware Parameter Estimation for Forecast Models in the … · 2015. 7. 29. · Context-Aware...

Context-Aware Parameter Estimation for Forecast Models in the Energy Domain

Lars Dannecker1,2, Robert Schulze1, Matthias Böhm2, Wolfgang Lehner2, Gregor Hackenbroich1

1SAP Research Dresden, 2Technische Universität Dresden

©  2011 SAP AG. All rights reserved. 2

Agenda

1.  Forecasting in the Energy Domain

2.  Context-Aware Forecast Model Repository

3.  Experimental Evaluation

4.  Summary and Future Work

Forecasting in the Energy Domain


Forecasting Process and Characteristics

Predicting the Future !  Quantitative model describing

historic time series behavior !  Uses parameters to represent

specific characteristic !  Estimated model mathematically

calculates future behavior

Specific Characteristics… …for energy time series •  Multi-Seasonality •  Dependence on external influences •  Evolving over time •  Negligible linear trend •  Continuous stream of

measurements

!

X t = "yt

It#S

+ 1#"( ) X t#1 + bt#1( )

bt = $ X t # X t#1( )+ 1# $( )bt#1

It = %Xt

X t+ 1# %( )It#S

Base Component

Trend Component

Season Component

ε ε


European Energy Market

Market Organizer

TSO TSO

BG2 BG3

Supply Demand

Balancing

Forecasting Aggregation

BG1

Balancing Energy Demand and Supply Guarantee stable grids !  Energy Demand has to be satisfied !  Penalties for oversupply !  Day-Ahead & intraday market !  Integration of more RES in power mix " Accurate predictions at any point in time

Renewable Energy Sources (RES) Increasing support !  Depending on uncertain influences !  Not plannable like traditional power " Accurate prediction for next day RES

supply necessary


Energy Data Management for Evolving Time Series

Energy Data Management Analytics close to the data !  Quick reactions to changing time series !  Always up-to-date forecasts

Appending new values over time !  Optimal parameters change and reoccur over time !  Multiple local minima in parameter space !  Continuous forecast model evaluation !  Efficient forecast model adaptation

New Measurements(consumption/

production)

Time Series

Forecast Model

( , , , state)

1

Model Update

Error Metric (SMAPE)

2 Model Evaluation Techniques

, , 3 Model Adaptation

Techniques

', , ,

Subsection 3.2 Subsection 3.3

Time Series

!!"#!"$

!"%

!"&###$#%#&

'!

'#

'$

'%

'&

$!

$#

$$

$%

!!"(!"#!"'!"$!")!"%!"*!"&!"+

,-./.

012,

3

,-425

!"

!#$"

!#%"

!#&"

!#'"

!#("

!#)"

!#*"

!#+"

!#,"

$"

%!!'-./0-%,"

%!!'-123-!'"

%!!(-456-$$"

%!!)-7/8-$+"

%!!*-9:;-%%"

%!!*-123-%,"

%!!+-<:=-!'"

%!!,-9:;-$,"

>/?/@:A:?"

">BC4<D" ">B7E1F4G" ">B<D>HD7CDI"

Context-Aware Forecast Model Repository


Context of Energy Time Series

Influences for Supply and Demand !  Time series development influenced

by background processes !  Changing context causes changes

demand and supply behavior !  Calendar: Special Days, Season !  Meteorological: Wind speed, Temp. !  Economical: Population

Context Drift Different types of drifting context

!

!

! !! !!

! ! ! !

! ! !!

!

! !!

!

!

! ! ! !

! ! !!

!

!

!

!

!

!

! ! ! !

! ! !


Case-Based Reasoning = Learning how to solve new problems from past experience Energy domain: Seasonal reoccurring contexts ! Reuse previous forecast models !  Retain: Save old parameter combinations with their respective context !  Retrieve: Search repository for a context most similar to the current context !  Revise: Use parameter combinations of similar context as input for optimization

Basic Idea

Problem-Solution Case Base

Start Values

Updating trigger

Continuous Insertions Continuous Forecasts

Updated Parameters

})({1 it pfz =+Time series

Current Forecast Model

3. Parameter Re-Estimation

Global Search Local Search

Forecast Error Calculation

Retrieve

Retain

Starting Values for Estimation

2. Parameter Storing and Retrieval

1. Model Evaluation

Model History Tree

}{ ip}{ ip}{ ipInsert

Retrieve

Distance Compuation

Revise Retain

Revise

Retrieve


Parameter Insertion

ContextSummary Parameters End Index

day hour year mean temperature p1 p2 K C1 3 5 2005 324 -12.3 0.684 0.34 16

C2 2 6 2006 648 -4.9 0.673 0.32 104

C3 5 5 2008 112 12.5 0.623 0.38 573

C4 4 8 2009 272 7.3 0.629 0.41 692

day

temperature

! 7 <7

<11.3

hour <10 ! 10

!11.3 <8 hour

! 8

year <2004 ! 2004

year <2005 ! 2005

mean <500 ! 500



C2 2 6 2006 648 -4.9 0.673 0.32 104

C3 5 5 2008 112 12.5 0.623 0.38 573

C4 4 8 2009 272 7.3 0.629 0.41 692

C5 4 1 2009 291 30.3 0.636 0.31 1024



C2 2 6 2006 648 -4.9 0.673 0.32 104

C3 5 5 2008 112 12.5 0.623 0.38 573

C4 4 8 2009 272 7.3 0.629 0.41 692

C5 4 1 2009 291 30.3 0.636 0.31 1024

PIQR 0.33 0.28 0.46 0.27 0.35



C2 2 6 2006 648 -4.9 0.673 0.32 104 ContextSummary Parameters End Index

day hour year mean temperature p1 p2 K C1 5 5 2008 112 12.5 0.623 0.38 573

C2 4 8 2009 272 7.3 0.629 0.41 692

C3 4 1 2009 291 30.3 0.636 0.31 1024

year <2008 ! 2008

1. Traverse to leaf node

2. Insert

4. Chose attribute with highest 5. Split

Tree Structured Repository !  Decision nodes: Splitting

attribute, splitting value !  Leaf nodes: Set of parameter

combinations, end index !  Splitting attributes chosen using

Partial Interquartil Range (PIQR) !  Split via partitioning median

3. " True/Split

!

c = 5 " 4 = cmax ?

!

PIQR =

˜ a 3 " ˜ a 1

2#

$ %

&

' (

aN " a1


Parameter Retrieval

x !"#"

$"

%"

&"

'"("

)"

J K

*"+"

,"

-"

."

/"

0"

1"

*2345674589:""+2345;97458:"

-2345<974569:"",2345;974599:"

/234569745;9:".23456=745<9:"

1. Traverse to corresponding leaf node

Find R as nearest neighbour

Find O as nearest neighbor Find P as nearest neighbor

!

a2

!

a1 = 0.65

!

a2 = 0.65

!

a1 = 0.9

!

a2 = 0.4

!

a1 = 0.35

!

a1 = 0.25

!"#"" $"%"&" '"(")" >"?" 1"0"

!

a1 = 0.7

4589" 456"

45@"

45A9"

45=9" 45A9"

!

a145<9"

2. Bob-test with " False " Ascent

!

a1 = 0.7

3. Bob-test with cyclical " True " Descent

!

a2 = 0.4

4. Bob-test with " False

!

a1 = 0.65

1

2

3 4


Optimization

Subsequence Similarity !  Find parameters that are associated with most similar time series shape !  Using Pearson Cross Correlation Coefficient

Subsequent Parallel Optimization !  Parallel local and global parameter optimization !  Local: Nelder Mead; Global: Simulated Annealing !  Results from local optimization directly used !  Parallel global search to consider areas not covered !  Global search continues after local search finished !  Quick accuracy recovery + global coverage

Current subsequence Old subsequence 2 Old subsequence 1

!

RZ " Z #( ) =zi $ z ( )

i=1

N $#

% " z i+# $ " z ( )&z2& " z

2

Local Optimization

ddrr

Global OptimizationParameterCandidates

CFMR

Experiments

NOTE: (Delete this element)

Sample of title slide image.

See SAP Image Library for other

available images.


Settings

DataSets !  UK National Grid: Aggregated Demand United Kingdom !  MeRegio: MeRegio project data 86 single customer demand !  NREL Wind: Aggregated data from US wind parks !  CRES PPV: Single appliance photovoltaic supply

Forecast Models !  Triple Seasonal Exponential Smoothing (5 parameters) !  EGRV multi-equation autoregressive model (up to 31 parameters

Comparison Scenario !  Time vs. Accuracy against 4 common approaches

Error Metric !  Symetric Mean Absolute Percentage Error (SMAPE)

Plattform !  AMD Athlon 4850e (2.5 GHz), 4GB RAM, Windows 7 !  Visual C++ 2010

Subsequent Parallel Optimization !  Parallel local and global parameter optimization !  Results from local optimization directly used !  Parallel global search to consider areas not covered !  Global search continues after local search finished !  Quick accuracy recovery + global coverage


Results: Triple Seasonal Exponential Smoothing

TS-Exponential Smoothing " Small number of parameters, quick to estimate !  MHT quickly reaches good accuracy !  Our method is not superior on all data sets !  Large result divergence for other approaches !  MHT overhead: Eval (100 models) " 4 msec, 20000 models " 0.6 sec


Results: EGRV Model (Energy Domain Specific)

EGRV " Large number of parameters, hard to estimate !  MHT achieved best results on all data sets !  Difference between best and worst approach much larger !  MHT better suited for more complex models !  MHT overhead: Eval (100 models) " 6 msec, 20000 models " 1.1 sec

Summary and Future Work


Summary & Future Work

Problem •  Evolving energy time series require efficient forecast model estimation

Summary •  Time series context influences time series development •  Case-based reasoning approach •  Store previous forecast model parameters for reuse with similar contextual situation

•  Tree organized Context-Aware Forecast Model Repository •  Retrieve parameter by comparing current context to past context •  Parameters serve as input for optimization approaches

Future Work •  Evaluate accuracy for approach without subsequent optimization • Order attributes in tree using information criterion •  Further parallelization

Thank You!

Contact information: Lars Dannecker SAP Research Dresden [email protected]


No part of this publication may be reproduced or transmitted in any form or for any purpose without the express permission of SAP AG. The information contained herein may be changed without prior notice. Some software products marketed by SAP AG and its distributors contain proprietary software components of other software vendors. Microsoft, Windows, Excel, Outlook, and PowerPoint are registered trademarks of Microsoft Corporation. IBM, DB2, DB2 Universal Database, System i, System i5, System p, System p5, System x, System z, System z10, System z9, z10, z9, iSeries, pSeries, xSeries, zSeries, eServer, z/VM, z/OS, i5/OS, S/390, OS/390, OS/400, AS/400, S/390 Parallel Enterprise Server, PowerVM, Power Architecture, POWER6+, POWER6, POWER5+, POWER5, POWER, OpenPower, PowerPC, BatchPipes, BladeCenter, System Storage, GPFS, HACMP, RETAIN, DB2 Connect, RACF, Redbooks, OS/2, Parallel Sysplex, MVS/ESA, AIX, Intelligent Miner, WebSphere, Netfinity, Tivoli and Informix are trademarks or registered trademarks of IBM Corporation. Linux is the registered trademark of Linus Torvalds in the U.S. and other countries. Adobe, the Adobe logo, Acrobat, PostScript, and Reader are either trademarks or registered trademarks of Adobe Systems Incorporated in the United States and/or other countries. Oracle is a registered trademark of Oracle Corporation. UNIX, X/Open, OSF/1, and Motif are registered trademarks of the Open Group. Citrix, ICA, Program Neighborhood, MetaFrame, WinFrame, VideoFrame, and MultiWin are trademarks or registered trademarks of Citrix Systems, Inc. HTML, XML, XHTML and W3C are trademarks or registered trademarks of W3C®, World Wide Web Consortium, Massachusetts Institute of Technology. Java is a registered trademark of Sun Microsystems, Inc.

JavaScript is a registered trademark of Sun Microsystems, Inc., used under license for technology invented and implemented by Netscape. SAP, R/3, SAP NetWeaver, Duet, PartnerEdge, ByDesign, SAP BusinessObjects Explorer, StreamWork, and other SAP products and services mentioned herein as well as their respective logos are trademarks or registered trademarks of SAP AG in Germany and other countries.

© 2011 SAP AG. All rights reserved

Business Objects and the Business Objects logo, BusinessObjects, Crystal Reports, Crystal Decisions, Web Intelligence, Xcelsius, and other Business Objects products and services mentioned herein as well as their respective logos are trademarks or registered trademarks of Business Objects Software Ltd. Business Objects is an SAP company.

Sybase and Adaptive Server, iAnywhere, Sybase 365, SQL Anywhere, and other Sybase products and services mentioned herein as well as their respective logos are trademarks or registered trademarks of Sybase, Inc. Sybase is an SAP company. All other product and service names mentioned are the trademarks of their respective companies. Data contained in this document serves informational purposes only. National product specifications may vary. The information in this document is proprietary to SAP. No part of this document may be reproduced, copied, or transmitted in any form or for any purpose without the express prior written permission of SAP AG. This document is a preliminary version and not subject to your license agreement or any other agreement with SAP. This document contains only intended strategies, developments, and functionalities of the SAP® product and is not intended to be binding upon SAP to any particular course of business, product strategy, and/or development. Please note that this document is subject to change and may be changed by SAP at any time without notice. SAP assumes no responsibility for errors or omissions in this document. SAP does not warrant the accuracy or completeness of the information, text, graphics, links, or other items contained within this material. This document is provided without a warranty of any kind, either express or implied, including but not limited to the implied warranties of merchantability, fitness for a particular purpose, or non-infringement. SAP shall have no liability for damages of any kind including without limitation direct, special, indirect, or consequential damages that may result from the use of these materials. This limitation shall not apply in cases of intent or gross negligence. The statutory liability for personal injury and defective products is not affected. SAP has no control over the information that you may access through the use of hot links contained in these materials and does not endorse your use of third-party Web pages nor provide any warranty whatsoever relating to third-party Web pages.

Context-Aware Parameter Estimation for Forecast Models in the … · 2015. 7. 29. · Context-Aware...

Documents

Transcript of Context-Aware Parameter Estimation for Forecast Models in the … · 2015. 7. 29. · Context-Aware...