Abstract
We consider interstage dependent stochastic linear programs where both the random right-hand side and the model of the underlying stochastic process have a special structure. Namely, for equality constraints (resp. inequality constraints) the right-hand side is an affine function (resp. a given function b t ) of the process value for the current time step t. As for m-th component of the process at time step t, it depends on previous values of the process through a function h tm .
For this type of problem, to obtain an approximate policy under some assumptions for functions b t and h tm , we detail a stochastic dual dynamic programming algorithm. Our analysis includes some enhancements of this algorithm such as the definition of a state vector of minimal size, the computation of feasibility cuts without the assumption of relatively complete recourse, as well as efficient formulas for sharing optimality and feasibility cuts between nodes of the same stage. The algorithm is given for both a non-risk-averse and a risk-averse model. We finally provide preliminary results comparing the performances of the recourse functions corresponding to these two models for a real-life application.
Similar content being viewed by others
Notes
If all lags p t (m) are null, we recover the case when process (ξ t ) is interstage independent.
The generalized autoregressive model is written using normalized random variables. For numerical reasons, it is recommended to use such formulation for the calibration of the model.
Exchanges between subsystems can also be considered. Demand satisfaction constraints can be written as INEQ in this case too.
Risk measures are defined on random variables representing costs, contrary to [4] where they are defined on random variables representing incomes. We easily switch from one setting to another since an income is the opposite of a cost.
We adopt the convention 1 \(\mathrm{MWMonth}=\frac{365.25\times24}{12}\) MWh=730.5 MWh.
The runs were done on a Dell PowerEdge 2900 server with 2 CPUs Intel Xeon E5345 (2.33 GHz, 8M of cache memory, 1333 MHz FSB), running under CentOS release 5, with 48 GB of RAM.
The computational time was approximately 4 weeks.
References
Birge, J.R.: Decomposition and partitioning methods for multistage stochastic linear programs. Oper. Res. 33, 989–1007 (1985)
Birge, J.R., Donohue, C.J.: The abridged nested decomposition method for multistage stochastic linear programs with relatively complete recourse. Algorithmic Oper. Res. 1, 20–30 (2001)
Birge, J., Louveaux, F.: Introduction to Stochastic Programming. Springer, New York (1997)
Guigues, V., Römisch, W.: Sampling-based decomposition methods for multistage stochastic programs based on extended polyhedral risk measures. SIAM J. Optim. 22, 286–312 (2012)
Guigues, V., Römisch, W.: SDDP for multistage stochastic linear programs based on spectral risk measures. Oper. Res. Lett. 40, 313–318 (2012)
Guigues, V., Sagastizábal, C.: Exploiting the structure of autoregressive processes in chance-constrained multistage stochastic linear programs. Oper. Res. Lett. 40, 478–483 (2012)
Guigues, V., Sagastizábal, C.: The value of rolling horizon policies for risk-averse hydro-thermal planning. Eur. J. Oper. Res. 217, 129–140 (2012)
Guigues, V., Sagastizábal, C.: Risk-averse feasible policies for large-scale multistage stochastic linear programs. Math. Program. 138, 167–198 (2013)
Homem de Mello, T., de Matos, V.L., Finardi, E.C.: Sampling strategies and stopping criteria for stochastic dual dynamic programming: a case study in long-term hydrothermal scheduling. Energy Syst. 2, 1–31 (2011)
Infanger, G., Morton, D.: Cut sharing for multistage stochastic linear programs with interstage dependency. Math. Program. 75, 241–256 (1996)
Maceira, M.E., Damázio, J.: The use of PAR(p) model in the stochastic dual dynamic programming optimization scheme used in the operation planning of the Brazilian hydropower system. In: VIII International Conference on Probabilistic Methods Applied to Power Systems, pp. 397–402 (2004)
Morton, D.P.: An enhanced decomposition algorithm for multistage stochastic hydroelectric scheduling. Ann. Oper. Res. 64, 211–235 (1996)
Pereira, M.V.F., Pinto, L.M.V.G.: Multi-stage stochastic optimization applied to energy planning. Math. Program. 52, 359–375 (1991)
Philpott, A., de Matos, V.: Dynamic sampling algorithms for multi-stage stochastic programs with risk aversion. Eur. J. Oper. Res. 218, 470–483 (2012)
Philpott, A.B., Guan, Z.: On the convergence of stochastic dual dynamic programming and related methods. Oper. Res. Lett. 36, 450–455 (2008)
Rockafellar, R.T., Uryasev, S.: Conditional value-at-risk for general loss distributions. J. Bank. Finance 26(7), 1443–1471 (2002)
Ruszczyński, A., Shapiro, A.: Conditional risk mappings. Math. Oper. Res. 31, 544–561 (2006)
Ruszczyński, A., Shapiro, A.: Optimization of convex risk functions. Math. Oper. Res. 31, 433–452 (2006)
Shapiro, A.: Analysis of stochastic dual dynamic programming method. Eur. J. Oper. Res. 209, 63–72 (2011)
Shapiro, A., Tekaya, W., da Costa, J.P., Soares, M.P.: Risk neutral and risk averse stochastic dual dynamic programming method. Eur. J. Oper. Res. 224, 375–391 (2013)
Wittrock, R.J.: Advances in a nested decomposition algorithm for solving staircase linear programs. Technical Report SOL 83-2, Systems Optimization Laboratory, Department of Operations Research, Stanford University, Stanford, CA (1983)
Acknowledgements
The author’s research was partially supported by an FGV grant, PRONEX-Optimization, and CNPq grant No. 382.851/07-4. We would like to thank the reviewers and the Associate Editor for beneficial comments and suggestions.
Author information
Authors and Affiliations
Corresponding author
Appendix
Appendix
Proof of Theorem 2.4
By duality, \(Q_{t}^{i} ( x_{t-1}, \xi_{[t-1]}, \eta_{t j} )\) may be expressed as the optimal value of the following linear program (due to Assumption (A3) the dual and the primal have the same finite optimal value):
where the objective function is given by
For problem (40), optimal solutions are extremal points of the feasible set. Further, the feasible set neither depends on x t−1 nor on ξ [t−1] and for any (x t−1,ξ [t−1]), row vectors \(\lambda_{t}^{k j}, \pi_{t 1}^{k j}, \pi_{t 2}^{k j}, \rho_{t}^{k j}\) are extremal points of the feasible set of problem \(Q_{t}^{i} ( x_{t-1}, \xi_{[t-1]}, \eta_{t j} )\) expressed as (40). It follows that \(Q_{t}^{i} ( x_{t-1}, \xi_{[t-1]}, \eta_{t j})\) is bounded from below by
for j=1,…,q t . Next, from the convexity of b ti , we obtain
and since \(\pi_{t 2}^{k j} \geq0\), we have
Plugging (44) into lower bound (42) for \(Q_{t}^{i} ( x_{t-1}, \xi_{[t-1]}, \eta_{t j})\) and since \(\mathcal{Q}_{t} ( x_{t-1}, \xi_{[t-1]})\) is bounded from below by \(\mathbb{E}_{\eta_{t}}[Q_{t}^{i} ( x_{t-1}, \xi_{[t-1]}, \eta_{t})] = \sum_{j=1}^{q_{t}} \, p(t,j) Q_{t}^{i} ( x_{t-1}, \xi_{[t-1]}, \eta_{t j})\), we obtain a cut of the form \(\theta_{t-1}^{k} + E_{t-1}^{k} x_{t-1} \geq {\tilde{E}}_{t-1}^{k} \xi_{[t-1]}+e_{t-1}^{k}\) and the result follows. □
Proof of Theorem 2.8
We show by induction, from t=T+1 down to t=2, that the announced cuts are valid and that \({\tilde{E}}_{t}^{k} \geq0\) for t=1,…,T, and k=0,1,…,iH. For t=T, we have \(\mathcal{Q}_{T+1}^{i}=\mathcal{Q}_{T+1} \equiv0\). As a result, all components of \(E_{T}^{k}, {\tilde{E}}_{T}^{k}\), and \(e_{T}^{k}\) are null for k=0,1,…,iH. In particular, we have that \({\tilde{E}}_{t}^{k} \geq0\). This achieves the first step of the induction.
Let us now assume that for some t∈{2,…,T}, valid cuts have been built for \(\mathcal{Q}_{\ell+1}, \ell=t,\ldots,T\), according to the formulas given in the theorem for \(E_{\ell}^{k}, {\tilde{E}}_{\ell}^{k}\), and \(e_{\ell}^{k}\), ℓ=t,…,T, k=0,1,…,iH, with all \({\tilde{E}}_{\ell}^{k} \geq0\).
We have \(\mathcal{Q}_{t}(x_{t-1}, \xi_{[t-1]})\) \(\geq\sum_{j=1}^{q_{t}} p(t,j) \mathcal{Q}_{t}^{i}(x_{t-1}, \xi_{[t-1]}, \eta_{t j})\) with
where
Let us first bound from below \(W_{t j}^{k}\). Using the convexity of b tℓ , we obtain
for every ℓ=1,…,ℓ t . Using these inequalities and the fact that \(\pi_{t 2}^{k j} \geq0\), we have
Similarly, using the convexity of h tm , we have
for every m=1,…,M. Using Assumption (A4), each component of each subgradient of b tℓ is nonnegative. As a result, all elements in matrix \(s_{t}^{b}( \xi_{t j}^{k})\) are nonnegative. Using this observation and relations (49) and (50), we obtain for \(W_{t j}^{k}\) the lower bound \(\pi_{t 2}^{k j} b_{t}( \xi _{t j}^{k})\) plus
Let us now bound from below \(V_{t j}^{k}\). Using relation (50) and the nonnegativeness of \({\tilde{E}}_{t, m}^{w}(1)\) (induction hypothesis) and of \(\rho_{t}^{k j}(w)\), we obtain for \(V_{t j}^{k}\) the lower bound
Plugging into (45) relation (46) as well as lower bounds (51) and (52) for respectively \(W_{t j}^{k}-\pi_{t 2}^{k j} b_{t}( \xi_{t j}^{k})\) and \(V_{t j}^{k}\), we obtain for \(\mathcal{Q}_{t}\) a cut of form \(-E_{t-1}^{k} x_{t-1}+{\tilde{E}}_{t-1}^{k} \xi_{[t]}+e_{t-1}^{k}\) with the desired values of \(E_{t-1}^{k}, {\tilde{E}}_{t-1}^{k}\), and \(e_{t-1}^{k}\). If remains to check that all components of \({\tilde{E}}_{t-1}^{k}\) are nonnegative. We had already observed that for all functions b ti , all components of all subgradients are nonnegative, due to Assumption (A4). The same remark holds for functions h tm , due to Assumption (A6). By induction hypothesis, all coefficients \(({\tilde{E}}_{t m}^{j}(\ell))_{j, \ell, m}\) are nonnegative. Using the nonnegativity of these coefficients, as well as the nonnegativity of row vectors \(\rho _{t}^{k j}\) and \(\pi_{t 2}^{k j}\), together with the formula for \({\tilde{E}}_{t-1}^{k}\), we obtain that \({\tilde{E}}_{t-1}^{k} \geq0,\;k=0,1,\ldots,iH\). □
Proof of Proposition 2.9
We show the cuts are valid when there are both equality and inequality constraints. A similar proof can be done when there are not equality constraints. Let k∈{(i−1)H+1,…,iH} and j∈{1,…,q t }. If \(\eta_{t j} \in\varOmega_{t}^{k}\) then \(Q_{t}^{i} ( x_{t-1}, \xi_{[t-1]}, \eta _{t j} )\) is bounded from below by \(g_{1}( \lambda_{t}^{k j}, \pi_{t 1}^{k j}, \pi_{t 2}^{k j}, \rho _{t}^{k j}, \eta_{t j} )\) where the expression of g 1 is given by (40). Next, for every j such that \(\eta_{t j} \notin\varOmega_{t}^{k}\), since all \((\lambda_{t}, \pi_{t 1}, \pi_{t 2}, \rho_{t}) \in\mathcal{M}_{t}^{i}\) belong to the feasible set of problem \(Q_{t}^{i} ( x_{t-1}, \xi_{[t-1]}, \eta_{t j} )\), we have \(Q_{t}^{i} ( x_{t-1}, \xi_{[t-1]}, \eta_{t j} ) \geq g_{1}( \lambda_{t}, \pi _{t 1}, \pi_{t 2}, \rho_{t}, \eta_{t j} )\) for every \((\lambda_{t}, \pi_{t 1}, \pi_{t 2}, \rho_{t}) \in\mathcal {M}_{t}^{i}\). As a result,
We then conclude as in the proof of Theorem 2.4. □
Proof of Theorem 2.10
Let x t−1 be a feasible state at the end of time step t−1 at a given node of this time step with history ξ [t−1]. Since for one of the son nodes, the realization of η t is η tj , the optimal value of (32) is 0. As a result, the optimal value of the dual of (32) is 0. This dual problem can be written
where the objective function f is given by
For this dual problem, since the optimal value is 0 and since \((\pi_{t 1}^{k j}, \pi_{t 2}^{k j}, \sigma_{t}^{k j})\) is feasible, we obtain
We conclude using (43) and (44). □
Proof of Proposition 2.11
We follow the proofs of Theorems 2.8 and 2.10. □
Proof of Theorem 3.1
The proof is similar to the proof of Theorem 2.4. □
Rights and permissions
About this article
Cite this article
Guigues, V. SDDP for some interstage dependent risk-averse problems and application to hydro-thermal planning. Comput Optim Appl 57, 167–203 (2014). https://doi.org/10.1007/s10589-013-9584-1
Received:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10589-013-9584-1