Stochastic Unit Commitment at ISO Scale: Issues and...
Transcript of Stochastic Unit Commitment at ISO Scale: Issues and...
![Page 1: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/1.jpg)
Stochastic Unit Commitment at ISO Scale: Issuesand Experiences
David L. Woodruff
University of California Davis
Paris, June 2014
1 / 29
![Page 2: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/2.jpg)
Outline
1 Overview
2 Some Unit Commitment Results
3 The Project
4 Progressive Hedging (PH)
5 Generating Scenarios
6 Conclusion
1 / 29
![Page 3: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/3.jpg)
The “Trailer”
Uncertainty in increasing due to renewables
Rather than simply increasing reserves, it may be possible toincorporate stochastics in Unit Commitment to save money
Solution methods are needed; we have them.
Scenarios with attached probabilities are needed; we have createdthem.
We are working on full, day-ahead, ISO-scale problems, but
we are not entering the debate on how to have a stochastic market.
So the immediate applications would be to reliability andadequacy as well as (mult-stage) planning.
2 / 29
![Page 4: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/4.jpg)
The “Trailer”
Uncertainty in increasing due to renewables
Rather than simply increasing reserves, it may be possible toincorporate stochastics in Unit Commitment to save money
Solution methods are needed; we have them.
Scenarios with attached probabilities are needed; we have createdthem.
We are working on full, day-ahead, ISO-scale problems, but
we are not entering the debate on how to have a stochastic market.
So the immediate applications would be to reliability andadequacy as well as (mult-stage) planning.
2 / 29
![Page 5: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/5.jpg)
The “Trailer”
Uncertainty in increasing due to renewables
Rather than simply increasing reserves, it may be possible toincorporate stochastics in Unit Commitment to save money
Solution methods are needed; we have them.
Scenarios with attached probabilities are needed; we have createdthem.
We are working on full, day-ahead, ISO-scale problems, but
we are not entering the debate on how to have a stochastic market.
So the immediate applications would be to reliability andadequacy as well as (mult-stage) planning.
2 / 29
![Page 6: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/6.jpg)
Project Logos
Sandia National Laboratories is a multi-program laboratory managed and operated by Sandia Corporation, a wholly owned subsidiary of Lockheed Martin Corporation, for
the U.S. Department of Energy’s National Nuclear Security Administration under contract DE-AC04-94AL85000.
Photos placed in horizontal position with even amount of white space
between photos and header
Photos placed in horizontal position with even amount of white space
between photos and header
To replace these boxes with images open
the slide master
To replace these boxes with images open
the slide master
3 / 29
![Page 7: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/7.jpg)
A Teaser
A few computational results.
I will leave out a lot of details.
We use Carrion and Arroyo as a staring point for the formulation
No network constraints
Validated against the Alstom test models
4 / 29
![Page 8: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/8.jpg)
WECC-240 Family
We use WECC-240-r1 (85 generators) for purposes of parametertuning and analysis.
Then fix the PH configuration and examine performance on theout-of-sample and more realistic WECC-240-r2 and WECC-240-r3cases.
We analyze scalability to the larger WECC-240-r2-x2 (170generators) and WECC-240-r2-x4 (340 generators) cases.
Using modest scale parallelism.
(These are harder to solve than ISO-NE instances of similar size.)
5 / 29
![Page 9: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/9.jpg)
Extensive form
Solution quality statistics for the extensive form of the WECC-240-r1instance, given 2 hours of run time.
Scenarios Obj Value MIP LB Gap % Run Time (s)
3 64279.18 63708.67 0.89 72915 62857.52 62052.75 1.26 730910 61873.01 60769.78 1.77 744425 61496.24 59900.40 2.59 773950 61911.74 59432.08 4.01 8279100 62388.85 3500.70 94.39 9379
6 / 29
![Page 10: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/10.jpg)
Larger Instances using PH
Solve time (in seconds) and solution quality statistics for PH executingon 50-scenario instances.
Instance Convergence Obj. Value PH L.B. Time
Red Sky Results
WECC-240-r2-x2 0.0 (in 22 iters) 117794.429 116538.868 741WECC-240-r2-x4 0.0 (in 19 iters) 232189.338 228992.984 1421
7 / 29
![Page 11: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/11.jpg)
Project Overview
Partners: Alstom Grid, Iowa State University, ISO-New England,Sandia National Laboratories, and the University of CaliforniaDavis
Funded by: US Dept. of Energy, ARPAe
Goal: Develop methods and test for solving stochastic unitcommitment (UC) problems at ISO scale including specification ofscenarios for uncertain demand and high penetration wind power.
We are estimating what energy savings would have occurred hadISO-NE been using stochastic UC for day-ahead commitmentduring the year 2011, but with high wind penetration.
8 / 29
![Page 12: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/12.jpg)
Day-ahead Stochastic UC
Unit commitments represent first stage decisions, which must bedetermined before values of uncertain parameters are known:
I intermittent resources,I loadsI unscheduled outages
Economic dispatch decisions provide recourse in the second stage,following resolution of parameter uncertainty.
9 / 29
![Page 13: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/13.jpg)
Scenario Trees
Scenario trees are used to represent the possible combinations ofparameter values (scenarios), in conjunction with theirprobabilities of realization.
Advanced scenario creation techniques have been leveraged tocompute cost confidence intervals and to minimize the number ofscenarios required to achieve reliable solutions.
10 / 29
![Page 14: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/14.jpg)
Scenario Trees
Scenario trees are used to represent the possible combinations ofparameter values (scenarios), in conjunction with theirprobabilities of realization.
Advanced scenario creation techniques have been leveraged tocompute cost confidence intervals and to minimize the number ofscenarios required to achieve reliable solutions.
10 / 29
![Page 15: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/15.jpg)
Solving with PH
Advanced decomposition strategies have been leveraged to meetrun-time and memory requirements, by distributing sub-problemsolutions across distinct computer nodes.
Progressive Hedging is a decomposition scheme that works wellwith mixed integer variables and specific non-linear structures,and is naturally parallelizable.
11 / 29
![Page 16: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/16.jpg)
Solving with PH
Advanced decomposition strategies have been leveraged to meetrun-time and memory requirements, by distributing sub-problemsolutions across distinct computer nodes.
Progressive Hedging is a decomposition scheme that works wellwith mixed integer variables and specific non-linear structures,and is naturally parallelizable.
11 / 29
![Page 17: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/17.jpg)
Progressive Hedging (PH)
Rockafellar and Wets (1991)
1 Decompose by scenario (a specification of the data with anattached probability)
2 Important: the expected value data can be one of the scenarios, sousers of PH will be no worse off than without PH
3 The algorithm parallelizes
12 / 29
![Page 18: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/18.jpg)
PH in Words
1 Solve every scenario as if it were the only problem
2 Form a non-aniticipative (but probably infeasible) solution byaveraging across the scenarios at nodes of the scenario tree
3 Compute multipliers (penalties or incentives) for each variable andeach scenario at non-leaf nodes
4 Solver every scenario, augmented by the penalities
5 If the scenarios solutions are similar enough, stop; otherwise gotostep 2.
13 / 29
![Page 19: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/19.jpg)
PH Diagram
PH Iteration 0:Solve Individual Scenario MIPs
Initialize Ws
Update Ws
Fix Variables That Have Converged
PH Iteration i:Solve Weighted Scenario MIPs
Progressive Hedging (PH)
Global Convergence Criterion Achieved?
“Done”
Standard MIP Solves
Key parameter:
Scenario-Based Decomposition
)( xxww xx
?|)(| xx
)( xxwx
2||||2/)(min xxxwxf x
14 / 29
![Page 20: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/20.jpg)
Example with 27 Scenarios(Probabilities not shown)
15 / 29
![Page 21: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/21.jpg)
Another Example with 27 Scenarios
16 / 29
![Page 22: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/22.jpg)
Yet Another Example with 27 Scenarios
17 / 29
![Page 23: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/23.jpg)
Last Example with 27 Scenarios
18 / 29
![Page 24: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/24.jpg)
Generating Scenarios: Overview
Most forecasts are given as a single (vector) point
To generate scenarios we need either:1 Good probabilistic forecasts, or2 Analysis of the error distributions of point forecasts.
A little notation:I The value (perhaps vector) of interest `I Leading indicator (if there is one) wI Forecast function (if there is one) `(w)
19 / 29
![Page 25: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/25.jpg)
Probabilistic Forecasts
One way: fit a function `(w), then find a way to generate wforecasts (or values) with known probabilities
Another way: have multiple forecast functions and assignprobabilities to each.
20 / 29
![Page 26: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/26.jpg)
Analysis of Error Distributions
You need, of course, a history of forecasts (or leading indicators)and corresponding observations.
It makes sense to group “similar conditions” thereby creating errordistributions that are conditional on the grouping.
If you are fitting your own forecast function, you can also segmentthe data based on forecast error characteristics and fit forecastthat are conditional on the error category.
21 / 29
![Page 27: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/27.jpg)
Similar Conditions
E.g.,
Forecast level quantiles
Derivative patterns
Weather quantiles
22 / 29
![Page 28: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/28.jpg)
Scenarios from Error Distributions
Using cutting points of the distribution to get skeleton points forthe scenarios from the center of the error distribution between thecutting points.
The probability of the skeleton points is trivial to compute fromthe cutting points.
Scenarios are formed by connecting skeleton points, if necessary.
23 / 29
![Page 29: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/29.jpg)
Scenario generationExample
24 / 29
![Page 30: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/30.jpg)
Scenario generationExample
25 / 29
![Page 31: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/31.jpg)
Conclusion
Due to renewables, uncertainty may need to be handled explicity,rather than via reserves.
To use Stochastic Programming, one needs algorithms andscenario generation methods
We have proposed both
Immediate application is most likely to be to intra-day reliabilitySCUC
Research continues and we have some data needs/requests...
26 / 29
![Page 32: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/32.jpg)
Rolling Horizon
Notre projet s’interesse a la facon dont les previsions de vents’ameliorent au cours du temps. Nous cherchons a obtenir des donneesindiquant :
les previsions de vent a differentes echeances (par exemple :prevision 2 jours avant, 1 jour avant, 6 heures avant...)
la quantite de vent reelle pour les dates correspondant auxprevisions. Elle nous permet de connaitre les erreurs de prediction.
27 / 29
![Page 33: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/33.jpg)
Continued...
Nous souhaitons avoir des predictions a differentes echeances pouretudier la facon dont les erreurs de prediction s’ameliorent au cours dutemps (plus on est proche de la date de prediction, plus l’erreur deprediction est faible). En etudiant la facon dont les predictionss’actualisent au cours du temps, nous souhaitons ainsi mettre en placeune methode pour mettre a jour les nouvelles predictions de vent etavoir une erreur de prediction plus faible.
28 / 29
![Page 34: Stochastic Unit Commitment at ISO Scale: Issues and ...cermics.enpc.fr/~delara/SESO/SESO2014/SESO2014-Thursday...Progressive Hedging (PH) Rockafellar and Wets (1991) 1 Decompose by](https://reader033.fdocuments.in/reader033/viewer/2022060921/60acc4581730be40ae67f93d/html5/thumbnails/34.jpg)
Exemple
Voici un exemple des donnees que nous souhaitons avoir :
les donnees reelles de vent pour chaque heure de la journee J (laperiode de temps peut varier : ca peut etre toutes les 5 minutes ouau contraire toutes les 2 heures)
les previsions de vent pour chaque heure de la journee J faites aujour J-1 a l’heure h (par exemple a 11h)
les previsions de vent pour chaque heure de la journee J faites aujour J-1 a l’heure h′ > h (par exemple a 23h)
Il s’agit d’un exemple. Nous sommes preneurs de tout type de donneesdu moment qu’elles concernent des previsions de vent et qu’ellespermettent d’etudier l’evolution des erreurs de prediction au cours dutemps.
29 / 29