A Hybrid Data Filtering Statistical Modeling Framework for Near-Term Forecasting

Transcription

1 A Hybrid Daa Filering Saisical Modeling Framework for Near-Term Forecasing Frank A. Monfore, Ph.D. Iron s Forecasing Brown Bag Seminar January 5, 2008

2 Please Remember In order o help his session run smoohly, your phones are mued. If you would like o make he presenaion porion of he screen larger, press he expand buon on he oolbar. Press i again o reurn o regular window. If you need echnical assisance during he meeing, dial *0 and you will be conneced o a Premiere Conferencing echnician. If you need o give oher feedback o he presener during he meeing, such as, slow down or need o ge he preseners aenion for some oher reason, use he pull down menu in he seaing char and we will address i righ away. If you have general quesions regarding he presenaion, please ype your quesion in he Q&A box in he boom, righ corner. We will ry o answer as many quesions as we can.

3 2008 Brown Bag Seminars A Hybrid Daa Filering-Saisical Modeling Framework for Near-Term Forecasing - January 5, 2008 Comparison of Load Research Expansion Mehods - April 29, 2008 Approaches o Consrucing Forecas Bounds - Augus 26, 2008 Using Neural Neworks o Improve Regression Models - December 2, 2008 All a noon, Pacific Time All will be recorded and available for review afer he session.

4 Iron Forecasing Background Iron has been deploying forecasing soluions for over 25 years for a range of companies and governmen agencies. > Worldwide user base > Independen sysem and marke operaors > Regional ransmission sysem operaors > Reailers operaing in one or more marke regions > Wholesalers, municipaliies and cooperaives > G&T uiliies, municipaliies, and cooperaives

5 A Hybrid Daa Filering-Saisical Modeling Framework for Near-Term Forecasing A Hich Hikers Guide o he Fuure

6 Conribuors The Model Framework presened here is based on a collaboraive effor. Special hanks go o: > Arhur Maniaci of he New York ISO, > Sen Li of he Midwes ISO, > Mark Quan, David Fabiszak, Suar McMenamin, and Jeff Fordham of Iron.

7 Scope of Near-erm Forecasing Problem Load forecas a he 5-minue level of load resoluion from 5 minues ahead ou 2 hours ahead Forecas is o be updaed every 5 minues leveraging off real-ime SCADA meering Wih marke resrucuring nodal level forecass are required Time is of he Essence

8 Forecas Challenges Volailiy in he 5 Minue SCADA Reads Hiing he Ramp Raes Hiing he Turning Poins (i.e. roughs, peaks) How do you decide how much useful informaion is in he mos recen 5 minue read? How useful is he mos recen daa read ou 5 minues, 5 minues, 30 minues, 60 minues or more?

9 6:20 PM 7:00 PM 2:20 PM 3:00 PM 3:40 PM 4:20 PM 5:00 PM 5:40 PM Forecasing he Morning Peak 4,000 3,750 3,500 3,250 3,000 2,750 2,500 :00 PM :40 PM 2:20 AM :00 AM :40 AM 2:20 AM 3:00 AM 3:40 AM 4:20 AM 5:00 AM 5:40 AM 6:20 AM 7:00 AM 7:40 AM 8:20 AM 9:00 AM 9:40 AM 0:20 AM :00 AM :40 AM 2:20 PM :00 PM :40 PM

10 7:00 7:00 PM PM 6:20 6:20 PM PM 5:00 5:00 PM PM 5:40 5:40 PM PM 4:20 4:20 PM PM 3:00 3:00 PM PM 3:40 3:40 PM PM 2:20 2:20 PM PM Forecasing he Morning Peak 4,000 3,750 3,500 3,250 3,000 2,750 2,500 :00 :00 PM PM :40 :40 PM PM 2:20 2:20 AM AM :00 :00 AM AM :40 :40 AM AM 2:20 2:20 AM AM 3:00 3:00 AM AM 3:40 3:40 AM AM 4:20 4:20 AM AM 5:00 5:00 AM AM 5:40 5:40 AM AM 6:20 6:20 AM AM 7:00 7:00 AM AM 7:40 7:40 AM AM 8:20 8:20 AM AM 9:00 9:00 AM AM 9:40 9:40 AM AM 0:20 0:20 AM AM :00 :00 AM AM :40 :40 AM AM 2:20 2:20 PM PM :00 :00 PM PM :40 :40 PM PM

11 7:00 PM 6:20 PM 5:40 PM 4:20 PM 5:00 PM 3:40 PM 3:00 PM Forecasing he Afernoon Peak 4,000 3,750 3,500 3,250 3,000 2,750 2,500 :00 PM :40 PM 2:20 AM :00 AM :40 AM 2:20 AM 3:00 AM 3:40 AM 4:20 AM 5:00 AM 5:40 AM 6:20 AM 7:00 AM 7:40 AM 8:20 AM 9:00 AM 9:40 AM 0:20 AM :00 AM :40 AM 2:20 PM :00 PM :40 PM 2:20 PM

12 7:00 PM 6:20 PM 5:00 PM 5:40 PM 3:40 PM 4:20 PM 3:00 PM Forecasing he Afernoon Peak 4,000 3,750 3,500 3,250 3,000 2,750 2,500 :00 PM :40 PM 2:20 AM :00 AM :40 AM 2:20 AM 3:00 AM 3:40 AM 4:20 AM 5:00 AM 5:40 AM 6:20 AM 7:00 AM 7:40 AM 8:20 AM 9:00 AM 9:40 AM 0:20 AM :00 AM :40 AM 2:20 PM :00 PM :40 PM 2:20 PM

13 Tradiional Approach: Daa Filering Mehods These mehods projec hisorical daa rends ino he fuure. > Kalman Filer Exponenial Smoohing Models > ARIMA Models Moving Average Models Polynomial Fis 5 Minue Acuals Daa Filer 5 Minue Forecas

14 ARIMA, MA, Polynomial Models ARIMA Model Load 0 + βload + β 2Load β j Load j β + ε Moving Average Model Load J j Load J j Polynomial Fi Load α + α + o α 2 + α 3 α 4

15 P + Q P Kalman Filer Updae Equaions ˆ ˆ L L P + P Q K Lˆ Measuremen Equaions P L P + + R K ( L L ) ˆ ˆ ( ) P P K

16 Kalman Filer Example () Updae Equaions Lˆ Q R P Q 0 R Measuremen Equaions Lˆ Lˆ Lˆ K Lˆ P + K P + R ( ˆ L L ) ( ) 00 P P 0 + Q P ( K ) P ( 0.5) 0 5

17 Kalman Filer Example (2) Updae Equaions Measuremen Equaions ˆ ˆ Q P P L L R P P K ( ) ( ) ( ) ( ) ˆ ˆ ˆ P K P L L K L L

18 Daa Filer Mehods Srenghs Weaknesses > Fas execuion, can be applied o a large number of nodes > Highly leverages he mos recen SCADA daa > Srong 5 o 0 minues ahead > Projecs mos recen rends ino he fuure, subjec o missing urning poins > Load volailiy will shadow iself ino he forecas Daa Filers are akin o driving a car while looking in he rear view mirror

19 Shadow Effec 3,500 3,450 3,400 3,350 3,300 3,250 3,200 3,50 3,

20 Shadow Effec 3,500 3,500 3,450 3,450 3,400 3,400 3,350 3,350 3,300 3,300 3,250 3,250 3,200 3,200 3,50 3,50 3, ,00

21 Saisical Model Approaches Generae load forecass given forecass of calendar, solar, and weaher condiions, as well as incorporae auoregressive erms. > Srucural Regression/Neural Nework Models wih or wihou auoregressive erms Calendar Weaher Solar Saisical Model 5 Minue Forecas 5 Minue Acuals

22 Saisical Model Approaches Srenghs > Forward looking, incorporaes fuure condiions (i.e. weaher, solar, calendar) > Can cach fuure urning poins/ramp raes > Can leverage he mos recen SCADA daa Saisical Models Map he Fuure Road Curvy Road Ahead

23 Saisical Model Approaches Srenghs > Forward looking, incorporaes fuure condiions (i.e. weaher, solar, calendar) > Can cach fuure urning poins/ramp raes > Can leverage he mos recen SCADA daa Weaknesses > Heavy IT fooprin making i difficul o generae muliple nodal forecass every 5 minues > If auoregressive erms are included hen subjec o he same problem wih load volailiy shadowing iself ino he forecas

24 A Hybrid Approach Solar? 5-Minue Forecas Calendar Weaher 5 Minue Acuals

25 Where We Were & Where We Are Going 4,00 4,000 Where We Were Where We Are Going 3,900 3,800 3,700 3,600 3,500 The Inen is o combine Daa on Where We Were wih Forecass Abou Where We Are Going 3,400 3,300 3,200 3,

26 Where We Are Going 5 Minue Acuals Aggregae o 5-Minue Loads Weaher Calendar Solar Saisical Model of 5 Minue Loads 5 Minue Forecas

27 Fi a Polynomial o Boh Informaion Ses 5 Minue Forecas 5 Minue Acuals 4,00 4,000 Where We Were Where We Are Going 3,900 Poly. 3,800 3,700 3,600 3,500 3,400 3,300 3,200 3, Polynomial Forecas

28 Where You Were & Where You Are Going 4,00 4,000 Where We Were Where We Are Going 3,900 Poly. 3,800 3,700 3,600 3,500 3,400 3,300 3,200 3,

29 Final Forecas Compared wih Acuals

30 Thank You. To know more, sar here:

31 Quesions? Press * o ask a quesion on he phone or ype in he box a he boom, righ corner. HANDS-ON WORKSHOPS Fundamenals of Sales and Demand Forecasing - March Orlando Forecasing 0 - April Washingon, DC Fundamenals of MerixND - May Washingon, DC Fundamenals of Shor-erm and Hourly Forecasing - Sepember Las Vegas Forecasing 0 - November San Diego USER MEETINGS European Forecasing User Meeing February - Brussels Ausralian Forecasing User Meeing - 2 March - Sydney Annual Energy Forecasing Group (EFG) Meeing - May Las Vegas 2nd Annual ISO/RTO Forecasing Summi - May Las Vegas Iron Users Conference - Ocober Dallas For more informaion and regisraion: Conac us a: , or forecasing@iron.com