A stochastic integer programming approach to the optimal thermal and wind generator scheduling problem

Transcription

1 A stochastic integer programming approach to the optimal thermal and wind generator scheduling problem Presented by Michael Chen York University Industrial Optimization Seminar Fields Institute for Research in Mathematical Science March 6th, 2012

2 Outline Overview

3 Outline Overview Problem statement

4 Outline Overview Problem statement Stochastic programming model

5 Outline Overview Problem statement Stochastic programming model Computational challenge and solution

6 Outline Overview Problem statement Stochastic programming model Computational challenge and solution Benchmarking report

7 Wind as a clean and renewable energy source

8 Current status of Canada wind energy

9 Legend Current status of Ontario wind energy ONTARIO WIND GENERATORS Greenwich 99 MW Bow Lake 20 MW Prince I 99 MW Goulais Wind Farm 25 MW Prince Il 90 MW McLean s Mountain 50 MW McLean s Mountain 10 MW Underwood (Enbridge) 181 MW Wolfe Island 198 MW Ripley South 76 MW Amaranth I 68 MW Farm-Owned Power 100 MW Amaranth II 132 MW Kingsbridge l 40 MW IN SERVICE IN DEVELOPMENT Conestogo 69 MW Map contains current and future grid-connected projects only Summerhaven 125 MW Port Dover and Nanticoke 105 MW Kruger Energy Chatham 99 MW Port Burwell (Erie Shores) 99 MW Pointe Aux Roches 49 MW Comber West 83 MW Gosfield 51 MW Spence Wind Farm (Talbot) 99 MW Dillon Wind Centre (Raleigh) 78 MW Port Alma (Kruger) 101 MW Comber East 83 MW January 2012

10 Can you balance? demand = generation?

11 Can you balance? demand = generation? stochastic demand = generation?

12 Can you balance? demand = generation? stochastic demand = generation? stochastic demand = stochastic generation?

13 be met by the sum of energy from various sources; (2) the inventory level in the Increasing device is balanced wind withand respect increasing to inflow and outflow system and energy volatility efficiency; (3) the pro level of each component of the system must be less than or equal to the capacity of th ponent. Tables 1 and 2 summarize the variables and parameters used in the model the objective. (a) Daily Demand Curves (b) Daily Wind Speed Curves Figure 2: Daily Curves for Demand and Wind Speed

14 Terrain Weather Prediction Wind Power Forecasting Demand Forecasting SCADA Deterministic Optimization System Monitoring Economic Dispatching Generator Status

15 All we need is prediction?

16 All we need is prediction? Prediction is very difficult, especially about the future. Niels Bohr

17 All we need is prediction? Prediction is very difficult, especially about the future. Niels Bohr Prediction + Operational Flexibility = Success!

18 All we need is prediction? Prediction is very difficult, especially about the future. Niels Bohr Prediction + Operational Flexibility = Success! we need more accurate forecasting;

19 All we need is prediction? Prediction is very difficult, especially about the future. Niels Bohr Prediction + Operational Flexibility = Success! we need more accurate forecasting;

20 All we need is prediction? Prediction is very difficult, especially about the future. Niels Bohr Prediction + Operational Flexibility = Success! we need more accurate forecasting; we need operational flexibility more than reserve can provide;

21 All we need is prediction? Prediction is very difficult, especially about the future. Niels Bohr Prediction + Operational Flexibility = Success! we need more accurate forecasting; we need operational flexibility more than reserve can provide;

22 All we need is prediction? Prediction is very difficult, especially about the future. Niels Bohr Prediction + Operational Flexibility = Success! we need more accurate forecasting; we need operational flexibility more than reserve can provide; we need operational flexibility less expensive than reserve.

23 Terrain Weather Prediction Wind Power Forecasting Demand Forecasting SCADA Stochastic Programming System Monitoring Robust Dispatching Generator Status

24 Benefit of stochastic programming approach 1. rigorous mathematical framework;

25 Benefit of stochastic programming approach 1. rigorous mathematical framework; 2. simutaneously considers many scenarios;

26 Benefit of stochastic programming approach 1. rigorous mathematical framework; 2. simutaneously considers many scenarios; 3. generates an optimal proactive baseline schedule; and a set of reactive adjustments;

27 Benefit of stochastic programming approach 1. rigorous mathematical framework; 2. simutaneously considers many scenarios; 3. generates an optimal proactive baseline schedule; and a set of reactive adjustments; 4. guranteed optimality;

28 Benefit of stochastic programming approach 1. rigorous mathematical framework; 2. simutaneously considers many scenarios; 3. generates an optimal proactive baseline schedule; and a set of reactive adjustments; 4. guranteed optimality; 5. proven confidence interval;

29 Benefit of stochastic programming approach 1. rigorous mathematical framework; 2. simutaneously considers many scenarios; 3. generates an optimal proactive baseline schedule; and a set of reactive adjustments; 4. guranteed optimality; 5. proven confidence interval; 6. more robust and secure grid;

30 Benefit of stochastic programming approach 1. rigorous mathematical framework; 2. simutaneously considers many scenarios; 3. generates an optimal proactive baseline schedule; and a set of reactive adjustments; 4. guranteed optimality; 5. proven confidence interval; 6. more robust and secure grid; 7. more integrated wind;

31 Benefit of stochastic programming approach 1. rigorous mathematical framework; 2. simutaneously considers many scenarios; 3. generates an optimal proactive baseline schedule; and a set of reactive adjustments; 4. guranteed optimality; 5. proven confidence interval; 6. more robust and secure grid; 7. more integrated wind; 8. lower energy cost for the society!

32 Stochastic programming model: key decision variables v ih : on/off status of slow generator i at each time h, i J s ;

33 Stochastic programming model: key decision variables v ih : on/off status of slow generator i at each time h, i J s ; v s ih : on/off status of fast generator i at each time h in scenario s, i J f, s S;

34 Stochastic programming model: key decision variables v ih : on/off status of slow generator i at each time h, i J s ; v s ih : on/off status of fast generator i at each time h in scenario s, i J f, s S;

35 Stochastic programming model: key decision variables v ih : on/off status of slow generator i at each time h, i J s ; v s ih : on/off status of fast generator i at each time h in scenario s, i J f, s S; All other variables are dependent on the variable v.

36 Stochastic programming model min first stage { }}{ j y jh + c d j z jh + c m j v jh ) + (cu h H j J s s.t. (y Js, z Js, v Js ) U Js second stage { }}{ π s( j y s jh + cd j z s jh + cm j v jh ) + c p ) j ps jh s S h H (cu h H j J f j J (1) (y Jf, z Jf, v Jf ) U s J f, s S (2) (p s, w s ) C s, s S (3) (p s, w s ) N s, s S (4) (p s, w s ) R s, s S (5)

37 Stochastic programming model: unit commitment constraints The unit commitment constraints for a generator j J is defined as: y jh v jh v j,h 1 h H (6a) U j := z jh v j,h 1 v jh h H (6b) y j,h Tj y jh v jh h H (6c) z j,h Tj z jh 1 v jh h H (6d) v j,h 1 v jh + y jh z jh = 0 h H (6e) v jh {0, 1}, y jh, z jh [0, 1] h H (6f) The constraint sets U J, U Js, U Jf are defined accordingly: U J := j J U j, U Js := j Js U j, U Jf := j Jf U j (7)

38 Stochastic programming model: reserve constraints R s := rs s jh η s dlh s h H (8a) j J l L rs s jh + ro s jh η dlh s h H (8b) j J j J l L

39 Stochastic programming model: reserve constraints R s := rs s jh η s dlh s h H (8a) j J l L rs s jh + ro s jh η dlh s h H (8b) j J j J l L We assume the contingency constraint is implemented exogenously following NERC N-1 rule.

40 Stochastic programming model: network constraints N s := p s jh + w s ih + fnmh s = j J m i I m nm E fmnh s + dlh s + gsh m m V, h H (9a) l L m mn E f s mnh = b mn(θ s mh θs nh γs mnh ) mn E, h H (9b) f mn fmnh s f mn mn E, h H (9c) γ mn γ s mnh γ mn mn E, h H (9d) = 0 h H (9e) θ s ref,h

41 Stochastic programming model: capacity constraints C s := P j v jh p s jh p s jh + rss jh P jv jh p s jh + rss jh + ros jh P j p s j,h 1 ps j,h R j p s jh ps j,h 1 + rss jh + ros jh R j j J, h H (10a) j J, h H (10b) j J, h H (10c) j J, h H (10d) j J, h H (10e) rs s jh RS j j J, h H (10f) ro s jh RO j w s lh ws lh j J, h H (10g) l L, h H (10h)

42 Our challenges: computational efficiency for stochastic linear programming problem, L-shape method is efficient and popular;

43 Our challenges: computational efficiency for stochastic linear programming problem, L-shape method is efficient and popular; stochastic integer programming problem is much more challenging!

44 Our challenges: computational efficiency for stochastic linear programming problem, L-shape method is efficient and popular; stochastic integer programming problem is much more challenging! both the first stage and the second stage of the stochastic unit commitment model are integer problem!

45 Our challenges: computational efficiency for stochastic linear programming problem, L-shape method is efficient and popular; stochastic integer programming problem is much more challenging! both the first stage and the second stage of the stochastic unit commitment model are integer problem! cutting-plane method is very effective for integer problem;

46 Our challenges: computational efficiency for stochastic linear programming problem, L-shape method is efficient and popular; stochastic integer programming problem is much more challenging! both the first stage and the second stage of the stochastic unit commitment model are integer problem! cutting-plane method is very effective for integer problem; how about stochastic integer problem?

47 Our solution: scenario crossing deep cuts Definition C sh J is a (s, h)-cover if P j + w s ih < (1 + η s) dlh s. j C sh i I l L If in addition P j + w s ih + P i (1 + η s ) dld s j C sh i I l L i J C sh, (11) then the cover C is simple.

48 Our solution: scenario crossing deep cuts PROPOSITION If (1 + η s ) l L ds 1 lh 1 i I ws 1 ih 1 (1 + η s ) l L ds 2 lh 2 i I ws 2 ih 2, then (i) a (s 1, h 1 )-cover is also a (s 2, h 2 )-cover; (ii) any (s 2, h 2 )-cover has a (s 1, h 1 )-cover subset.

49 Our solution: scenario crossing deep cuts PROPOSITION Let C be a (s, h)-cover and s h = l L ds lh j C P j i I ws ih. Then the strengthened (s, h)-cut j J C p s jh is valid for ρ s ( ). If in addition, max{p j, s h } + i I w s ih s h 1 (12) s h P j, j J C, and (11) holds strictly for some indices, then (12) is facet-defining for ρ s ( ).

50 Our solution: scenario crossing deep cuts Definition Two generators a and b are symmetric if a and b have identical physical features and are located on one bus. PROPOSITION Assume there are κ symmetric pairs in the electricity grid. Let Ω s be the feasible set of (v s, y s, z s ) in ρ s (λ s ), and Ω s be the reduced feasible set after applying the κ symmetry cuts: then y s ah + vs a,h 1 + vs b,h 1 ys bh, for all symmetric pairs (a,b), Ω s = Ωs 2 κ, i.e., the feasible region shrinks exponentially.

51 Numerical test: RTS-96 system Table: Generator Mix Type Technology No. units Capacity(MW) list of units U12 Oil/Steam U20 Oil/CT ,5-6 U50 Hydro U76 Coal/Steam , 7-8 U100 Oil/Steam U155 Coal/Steam , U197 Oil/Steam U350 Coal/3 Steam U400 Nuclear W150 Wind W100 Wind

52 Numerical test: RTS-96 system Table: Bus Generator Incidence Bus Generators Bus Generator Bus Generator W W

53 IEEE eering ry 21- made ore tion the the ata not ata led of t it of the ith ave any the TSges this ata. M. nty, M. 30 generators; 24 buses; 34 lines; installed capacity 3,405MW. 38 kv Figure 1 - IEEE One Area RTS-96

54 Effect of symmetry cut Optimal schedule 1 Optimal schedule 2 Gen switch on switch off Gen switch on switch off Table: Optimal Schedule for Two-scenario Instance. Generators not shown in the table are always on. No wind curtailment nor demand shedding appears in this optimal schedule. The optimal objective value is $798,256. Total demands for the two scenarios over the 24 hrs are MW and MW.

55 Benchmarking report CPLEX B&B CPLEX D& S Cover Cuts Both cuts S nodes seconds nodes seconds nodes seconds Ncuts node second Table: Statistics of running time. The number of cuts shown in table is for cover cuts; the number of symmetry cuts is constantly 36, which is not shown in the table. S mean median cover cut 29% 76% 24% 78% 17% 30% 42% 29% both cuts 50% 86% 85% 76% 93% 68% 76% 80% Table: Reduction rate of running time

56 Conclusion Canada could benefit from the abundant wind resource; if

57 Conclusion Canada could benefit from the abundant wind resource; if the system volatility could be safely controlled.

58 Conclusion Canada could benefit from the abundant wind resource; if the system volatility could be safely controlled. along with prediction, operational flexibility is an extremely important mechanism;

59 Conclusion Canada could benefit from the abundant wind resource; if the system volatility could be safely controlled. along with prediction, operational flexibility is an extremely important mechanism; Stochastic integer programming approach can greatly enhance the operational flexibility; however,

60 Conclusion Canada could benefit from the abundant wind resource; if the system volatility could be safely controlled. along with prediction, operational flexibility is an extremely important mechanism; Stochastic integer programming approach can greatly enhance the operational flexibility; however, computational efficiency is its bottleneck.

61 Conclusion Canada could benefit from the abundant wind resource; if the system volatility could be safely controlled. along with prediction, operational flexibility is an extremely important mechanism; Stochastic integer programming approach can greatly enhance the operational flexibility; however, computational efficiency is its bottleneck. We build a mathematically rigorous stochastic optimization model for the problem.

62 Conclusion Canada could benefit from the abundant wind resource; if the system volatility could be safely controlled. along with prediction, operational flexibility is an extremely important mechanism; Stochastic integer programming approach can greatly enhance the operational flexibility; however, computational efficiency is its bottleneck. We build a mathematically rigorous stochastic optimization model for the problem. Our research on cross-scenario deep cuts speeds up the state-of-art CPLEX solver significantly!

63 Future research: can stochastic programming approach safely reduce system reserve?

64 Future research: can stochastic programming approach safely reduce system reserve? how to model peaks in a 5-minutes interval efficiently?

65 Future research: can stochastic programming approach safely reduce system reserve? how to model peaks in a 5-minutes interval efficiently? what is an optimal generator mix for future market?

66 Future research: can stochastic programming approach safely reduce system reserve? how to model peaks in a 5-minutes interval efficiently? what is an optimal generator mix for future market? considering an ISO DACP algorithm, how to make best bids?

67 Future research: can stochastic programming approach safely reduce system reserve? how to model peaks in a 5-minutes interval efficiently? what is an optimal generator mix for future market? considering an ISO DACP algorithm, how to make best bids? how to evaluate the efficiency, reliability and profitability of an energy producer?

68 Future research: can stochastic programming approach safely reduce system reserve? how to model peaks in a 5-minutes interval efficiently? what is an optimal generator mix for future market? considering an ISO DACP algorithm, how to make best bids? how to evaluate the efficiency, reliability and profitability of an energy producer? are there benefit for optimally switching transmission line?

69 Future research: can stochastic programming approach safely reduce system reserve? how to model peaks in a 5-minutes interval efficiently? what is an optimal generator mix for future market? considering an ISO DACP algorithm, how to make best bids? how to evaluate the efficiency, reliability and profitability of an energy producer? are there benefit for optimally switching transmission line? how much gain a new transmission line can bring to a grid system?

70 Future research: can stochastic programming approach safely reduce system reserve? how to model peaks in a 5-minutes interval efficiently? what is an optimal generator mix for future market? considering an ISO DACP algorithm, how to make best bids? how to evaluate the efficiency, reliability and profitability of an energy producer? are there benefit for optimally switching transmission line? how much gain a new transmission line can bring to a grid system? how much gain a battery array can bring to a grid system?

71 Future research: can stochastic programming approach safely reduce system reserve? how to model peaks in a 5-minutes interval efficiently? what is an optimal generator mix for future market? considering an ISO DACP algorithm, how to make best bids? how to evaluate the efficiency, reliability and profitability of an energy producer? are there benefit for optimally switching transmission line? how much gain a new transmission line can bring to a grid system? how much gain a battery array can bring to a grid system? how to integrate wind forecasting and unit commitment?

72 Future research: can stochastic programming approach safely reduce system reserve? how to model peaks in a 5-minutes interval efficiently? what is an optimal generator mix for future market? considering an ISO DACP algorithm, how to make best bids? how to evaluate the efficiency, reliability and profitability of an energy producer? are there benefit for optimally switching transmission line? how much gain a new transmission line can bring to a grid system? how much gain a battery array can bring to a grid system? how to integrate wind forecasting and unit commitment? how to integrate demand management and production scheduling?

73 Paper and slides are available upon request, please contact:

74 Paper and slides are available upon request, please contact: Assistant Professor Michael Chen Mathematics and Statistics Department York University, Toronto ext