Robust Optimization for Unit Commitment and Dispatch in Energy Markets

Transcription

1 1/41 Robust Optimization for Unit Commitment and Dispatch in Energy Markets Marco Zugno, Juan Miguel Morales, Henrik Madsen (DTU Compute) and Antonio Conejo (Ohio State University) Modeling and Optimization of Heat and Power System DTU, 12th January 2015

2 2/41 Outline 1 Mathematical Framework 2 Energy & Reserve Dispatch 3 CHP Unit Commitment & Scheduling 4 Conclusion

3 3/41 Why Optimization Under Uncertainty Sources of uncertainty in energy systems Decision-making in energy markets Better decisions if optimization accounts for uncertainty

4 4/41 Deterministic Optimization Framework Unit commitment and dispatch formalized as optimization problems Min. x,y c x + q y s.t. Ax b, Tx + Wy h x, y represent decision variables: unit status, dispatch, storage level, network state, etc. inequality and equality constraints impose limits on dispatch: dispatch limits, ramping, supply/demand balance, etc.

5 5/41 Optimization Under Uncertainty Most often not all parameters are known in advance { } Min. c x + M ω q x,y ω ωy ω s.t. Ax b, Tx + Wy ω h ω, ω Ω Uncertain parameters depend on random variable ω Variables y ω are adjustable (recourse) Measure of recourse cost in the objective function (expectation, etc.) Constraints hold for different realizations of the uncertainty

6 6/41 Stochastic Programming Tractable version of stochastic problem by sampling the uncertainty Min. x,y ωs c x + S p ωs q ω s y ωs s=1 s.t. Ax b, Tx + Wy ωs h ωs, s = 1,..., S Probably the most popular method of optimization under uncertainty Only used as a comparison in this presentation

7 7/41 Adaptive Robust Optimization (ARO) Tractable version of stochastic problem based on notion of uncertainty set Min. x c x+ max s.t. Ax b, min q,h y q y Tx + Wy h, (q, h) U, M( ) replaced by worst-case realization of recourse cost Inner minimization problem guarantees feasibility of solution Three-level game against nature

8 8/41 ARO and Non-Anticipativity t 1 t 2 t 3 t 4 Worst-case trajectory of uncertainty revealed before making recourse decisions Recourse has perfect knowledge of future worst-case uncertainty

9 9/41 Affinely Adjustable Robust Optimization (AARO) Enforcing Linear Decision Rules Optimal recourse functions y(ω) approximated by linear rules: y = Yω Computational Simplification Linear coefficient Y is a first-stage variable: essentially a max-min problem y ω

10 10/41 Affinely Adjustable Robust Optimization (AARO) Problem Formulation { } Min. x,y c x + E ω ω Q Yω s.t. Ax b, Tx + WYω Hω, ω U { } Min. x,y c x + E ω ω Q Yω s.t. Ax b, max {Tx + (WY H) ω} 0 ω Ω Differences from ARO: Not bound to optimize worst-case cost: can be made less conservative min-max problem can be simplified via duality argument: single stage problem

11 11/41 Energy and Reserve Dispatch using Adaptive Robust Optimization

12 12/41 Electricity Market Setup Day-ahead market: 12- to 36-hour ahead Energy and reserve are co-optimized (first stage) Followed by a balancing (real-time) market (second stage) Network is represented as in US markets The right of starting-up and shutting-down own units is not confiscated as in EU markets

13 13/41 Adaptive Robust Optimization Approach Two-stage robust optimization turn into three-level optimization problems 1 Make dispatch decision x with prognosis of the future (minimize cost) 2 Uncertainty (wind power production w) unfolds (maximize cost) 3 Make operation (recourse) decision y (minimize cost)

14 14/41 The Robust Optimization Approach min x c x T x +max w s.t. Fx = d ŵ, Gx g. min y c y T y s.t. Py = w Qx, : λ, s.t. H w h, Ly l Mx N w, : µ, This is a complicated problem! There is no general solution technique for three-level optimization problems

15 15/41 Modeling the Uncertainty Set Uncertain stochastic power production at different nodes described by polyhedral uncertainty set W can be modeled via linear inequalities H w h can include budget constraints on the stochastic production can limit the variation of stochastic production among adjacent plants w 1 w 2 ŵ W

16 Column & Constraint Generation Approach In the linear case, for any feasible x the worst-case realization of w is a vertex of the polytope W the worst-case recourse cost is the maximum of a finite number of affine functions of x 1 Find optimal first-stage solution for approximation (lower bound) 2 Calculate worst-case recourse cost (upper bound) 3 Generate cut 16/41

17 Column & Constraint Generation Approach In the linear case, for any feasible x the worst-case realization of w is a vertex of the polytope W the worst-case recourse cost is the maximum of a finite number of affine functions of x Cost 1 Find optimal first-stage solution for approximation (lower bound) 2 Calculate worst-case recourse cost (upper bound) 3 Generate cut (p,r) 16/41

19 Column & Constraint Generation Approach In the linear case, for any feasible x the worst-case realization of w is a vertex of the polytope W the worst-case recourse cost is the maximum of a finite number of affine functions of x Cost gap 1 Find optimal first-stage solution for approximation (lower bound) 2 Calculate worst-case recourse cost (upper bound) 3 Generate cut (p,r) 16/41

22 Column & Constraint Generation Approach In the linear case, for any feasible x the worst-case realization of w is a vertex of the polytope W the worst-case recourse cost is the maximum of a finite number of affine functions of x 1 Find optimal first-stage solution for approximation (lower bound) 2 Calculate worst-case recourse cost (upper bound) 3 Generate cut Cost gap (p,r) 16/41

23 Column & Constraint Generation Approach In the linear case, for any feasible x the worst-case realization of w is a vertex of the polytope W the worst-case recourse cost is the maximum of a finite number of affine functions of x 1 Find optimal first-stage solution for approximation (lower bound) 2 Calculate worst-case recourse cost (upper bound) 3 Generate cut Cost gap (p,r) 16/41

25 17/41 Modified IEEE 24-Node Reliability Test System Block offers by generators Flexible generators provide reserve Three lines with reduced capacity Wind farms at 6 nodes of the system

26 18/41 Wind Power Data Data from 6 locations in Spain Uncertainty set parameters estimated from historical data Historical data also used as scenarios

27 19/41 Uncertainty Set for Wind Power Production Budget of uncertainty: Intervals: Wn max Budget: w n W max n Φ N n w n W max n, n Φ N Γ One additional type of constraint to model geographical correlation ρ n1 n 2 w n 1 w n 2 ρ wn max 1 wn max n1 n 2, n 1, n 2 Φ N, n 1 n 2 2

28 Robust Optimization vs Stochastic Programming Risk-Neutral Case Cost Robust Optimization Stochastic Programming Energy disp Up reserve Down reserve Total day-ahead Mean CVaR 95% VaR 100% Mean CVaR 95% VaR 100% Energy redisp Load-shedding Total balancing Total aggregate /41

29 21/41 Robust Optimization vs Stochastic Programming Risk-Averse Case CVaR Expectation Cost [$] Cost [$] 2.3 x In-sample α 2.9 x SP RO SP RO Cost [$] Cost [$] 2.35 x Out-of-sample α 2.95 x SP RO SP RO α α

30 22/41 Robust Optimization vs Stochastic Programming Computational Properties Computational results with different IEEE power system models Computation time (s) 10 2 SP (24 bus) 10 1 RO (24 bus) SP (118 bus) RO (118 bus) SP (300 bus) 10 0 RO (300 bus) # scenarios

31 22/41 Unit Commitment and Scheduling for Heat & Power Systems using Affinely Adjustable Robust Optimization

32 23/41 Heat and Power System Example Heat plant District heating CHP (extraction) CHP (backpressure) Heat tank Electricity market

33 24/41 A Multi-Stage Uncertain Problem The problem is uncertain (heat demand, prices) and multi-stage: Day-ahead Determine unit commitment and schedule for the following day (periods 1,..., N T ) Oper. h 1 Uncertainty at time 1 unfolds and units are re-dispatched (based on realizations until time 1) Oper. h 2 Uncertainty at time 2 unfolds and units are re-dispatched (based on realizations until time 2). Oper. h N T Uncertainty at time N T unfolds and units are re-dispatched (based on all realizations)

37 25/41 Robust Optimization Approach Our approach based on Robust Optimization (RO) } Minimize c x + E ω {q ωy(ω) x,y( ) where s.t. Ax b, Tx + Wy(ω) h ω, ω U (3) x are first-stage decision (unit commitment, schedule) y( ) are recourse functions of the uncertainty ω (unit operation) The solution is optimal in expectation (1) The solution is feasible for any ω in an uncertainty set U (3) (1) (2)

38 26/41 Linear Decision Rules Optimal recourse decision approximated via linear decision rules y t y it (ω) = ŷ it + Y itτ ω τ, τ=1 i, t ω Recourse decision depends on history of uncertainty up to now (t) Independence on future uncertainty realizations (τ > t): non-anticipative

39 27/41 Final Linear Reformulation We make the following assumptions: linear dependence of parameters to uncertainty: q ω = Qω, h ω = Hω budget uncertainty set defined by Dω d By duality arguments, the problem can be reformulated linearly: { Min. x,y,λ c x + tr Q Y (Σ ω + E{ω}E{ω} )} s.t. Ax b, Λ d Tx, D Λ = (WY H), Λ R R L 2 0

40 28/41 Case-Study Setup Three production facilities with storage Extraction CHP (AMV3, AVV1) Back-pressure CHP (AMV1) heat-only plant Historical heat load from VEKS RO: budget uncertainty set (tunable interval size and budget) SP: Gaussian scenarios Power prices from NordPool SP: Gaussian scenarios correlated with heat demand Figure from HOFOR.dk

41 29/41 Back-pressure vs Extraction CHP Units P Back-pressure CHP unit constant heat/power ratio (r b ) rb H Extraction CHP unit more flexible operation in heat/power space feasible region approximated by polyhedron P rv rb H

42 Parameters of the System Extraction Back-pressure Heat-only Power/heat ratio r b Electricity loss for heat prod. r v Fuel per electricity unit ϕ p Fuel per heat unit ϕ q Minimum fuel input f Maximum fuel input f Ramp-up/-down for fuel input r/ r Minimum heat output q Maximum heat output q Fuel marginal cost c No-load cost c Start-up/shut-down cost c SU /c SD Minimum up-time T U Minimum down-time T D Provides real-time flexibility yes no yes 30/41

43 31/41 Scheduling Example Results from the Deterministic Model Time period (h) 600 Extraction Backpressure Storage Demand Schedule (MWh) Time period (h)

44 32/41 Scheduling Example Results from the Robust Optimization Model Time period (h) 600 Extraction Backpressure Storage Demand Schedule (MWh) Time period (h)

45 Tuning the Need for Flexibility The extraction unit is the main provider of flexibility Unit Γ Total heat dispatch MWh Periods online h /41

46 34/41 Piecewise Linear Decision Rules We can define different decision rules for partitions of the uncertainty set 2 different rules for each dimension of the uncertainty partitioning demand deviation about 0 straightforward choice different response to deficit/surplus heat demand y LDR PLDR ω

47 Linear vs Piecewise Linear Decision Rules Dispatch Improvement Dispatch for extraction unit during peak-demand day Decision rule Unit Γ Linear MWh Piecewise-linear MWh /41

48 Linear vs Piecewise Linear Decision Rules In-Sample vs Out-of-Sample Results Theoretical improvement Empirical improvement % Γ Γ % Interval size (# sd) Interval size (# sd) decision rules not binding in practice anticipative out-of-sample evaluation (stochastic programming) 36/41

49 Price of Robustness Sensitivity analysis can help tune the RO parameters (cost vs robustness) Average profit Worst-case load-not-served (MWh) Γ Γ Interval size (# sd) Interval size (# sd) 37/41

50 RO vs Alternative Methods Out-of-Sample Analysis Model Interval Load not served (MWh) Γ Average profit (e) radius (# sd) largest expected RO-PLDR , DET SP Load-not-served not penalized in the objective function Out-of-sample analysis based on SP 38/41

51 39/41 Conclusion Robust optimization addresses some of the shortcomings of stochastic programming intractability (large scenario sets) violation of non-anticipativity in multi-stage problems Simpler models of the uncertainty are required uncertainty sets low-order moments (mean, standard deviation, etc.) Trade-off between optimality and conservativeness can be tuned via uncertainty set parameters similar results in terms of CVaR for risk-averse SP (ARO model for energy and reserve) tunable conservativeness in AARO model for CHP units Promising results in terms of scalability

52 40/41 References M. Zugno, A.J. Conejo (2014). A Robust Optimization Approach to Energy and Reserve Dispatch in Electricity Markets, (under review). M. Zugno, J.M. Morales, H. Madsen (2014). Commitment and Dispatch of Heat and Power Units via Affinely Adjustable Robust Optimization (working paper, available soon). M. Zugno, J.M. Morales, H. Madsen (2014). Robust Management of Combined Heat and Power Systems via Linear Decision Rules, 2014 IEEE Energy Conference (ENERGYCON) Proceedings,

53 41/41 Thank you for your attention! Q&A Session

54 42/41 Inner Problem Reformulation via Duality From min-max-min to min-max-max: min x c x T x+max w s.t. Fx = d ŵ, Gx g. max λ,µ ( w Qx)T λ + (l Mx N w) T µ s.t. P T λ + L T µ = c y, µ 0, s.t. H w h, Inner max-max problem can be merged into a single bilinear maximization problem

55 43/41 Solution via Cutting-Plane Method min x,β c x T x + β s.t. β w T k λ k + (l N w k ) T µ k (λ k Q + µ k M) x, k, Fx = d ŵ, Gx g. 1 Start with a lower bound β for the recourse cost 2 Find optimal first-stage solution x for the approximation (lower bound) 3 Calculate worst-case recourse cost (upper bound) 4 Generate cut for the vertex, go back to 2

56 43/41 Solution via Cutting-Plane Method min x,β c x T x + β s.t. β w T k λ k + (l N w k ) T µ k (λ k Q + µ k M) x, k, Fx = d ŵ, Gx g. While (upper bound - lower bound > ɛ) 1 Start with a lower bound β for the recourse cost 2 Find optimal first-stage solution x for the approximation (lower bound) 3 Calculate worst-case recourse cost (upper bound) 4 Generate cut for the vertex, go back to 2

60 So far only easy-to-enumerate polyhedral sets have been considered, or heuristic techniques 44/41 Subproblem Choice vs Uncertainty Set Point 3 in the cutting-plane method requires the solution of a bilinear program Simple, easy-to-enumerate uncertainty sets (budget) Heuristic methods (suboptimal) We want to be able to solve the problem exactly for any polyhedral uncertainty set w 2 ŵ W w 1

61 45/41 Reformulation of the Problem (1) Swapping Inner Problems The inner problems can be swapped: min x c x T x+max λ,µ s.t. Fx = d ŵ, Gx g, (Qx)T λ + (l Mx) T µ+max w s.t. P T λ + L T µ = c y, µ 0, ( λ T + µ T N ) w s.t. H w h, : ξ,

62 46/41 Reformulation of the Problem (2) Using Strong Duality The inner problem is linear strong duality holds: ( λ T + µ T N ) w = h T ξ } {{ } objective function We can exchange a bilinear term with a linear one, but the innermost problem becomes a minimization problem We can enforce KKT conditions to guarantee optimality and get rid of a level

63 indicates that the product of the operands is 0 this can be linearized using binary variables (MILP) 47/41 Reformulation of the Problem (3) Final Formulation min x c x T x+ max λ,µ, w,ξ (Qx)T λ + (l Mx) T µ + h T ξ s.t. Fx = d ŵ, Gx g. s.t. 0 ξ h H w 0, H T ξ = λ N T µ, P T λ + L T µ = c y, µ 0,

64 RO vs Alternative Methods Out-of-Sample Analysis (Spring) Model Interval radius (# sd) Γ Average profit (e) RO-PLDR Heat load not served (MWh) largest expected DET SP /41

65 RO vs Alternative Methods Out-of-Sample Analysis (Summer) Model Interval radius (# sd) Γ Average profit (e) Heat load not served (MWh) largest expected RO-PLDR DET SP /41

66 RO vs Alternative Methods Out-of-Sample Analysis (Autumn) Model Interval radius (# sd) Γ Average profit (e) RO-PLDR Heat load not served (MWh) largest expected DET SP /41