Stochastic Lagrangian Relaxation in Power Scheduling of a Hydro-Thermal System under Uncertainty

We consider a power generation system comprising thermal units and pumped hydro storage plants, and introduce a model for its weekly cost-optimal operation. Due to the uncertainty of the load, the mathematical model represents a dynamic (multi-stage) stochastic program. The model involves a large number of mixed-integer (stochastic) decisions but its constraints are loosely coupled across operating power units. The coupling structure is used to design a stochastic Lagrangian relaxation method, which leads to a decomposition into stochastic single unit subproblems. The stochastic subproblems have deterministic counterparts, which makes it easy to develop algorithms for the stochastic problems. In this paper, a descent method for stochastic storage problems and an extension of dynamic programming towards stochastic programs are developed. The solution of the dual problem provides multipliers leading to preferred schedules (binary primal variables). The crossover heuristics evaluates the economic dispatch problems corresponding to a sequence of such preferred schedules. The combination of the restriction on dual preferred schedules (Lagrangian reduction) with the evaluation of a sequence (facet search) leads to an eÆcient method. The numerical results on realistic data of a German utility justify this approach.