Stochastic optimization of unit commitment: A new decomposition framework

This paper presents a new stochastic decomposition method well-suited to deal with large-scale unit commitment problems. In this approach, random disturbances are modeled as scenario trees. Optimization consists in minimizing the average generation cost over this ''tree-shaped future''. An augmented Lagrangian technique is applied to this problem. At each iteration, nonseparable terms introduced by the augmentation are linearized so as to obtain a decomposition algorithm. This algorithm may he considered as a generalization of price decomposition methods, which are now classical in this field, to the stochastic framework. At each iteration, for each unit, a stochastic dynamic subproblem has to be solved. Prices attached to nodes of the scenario trees are updated by the coordination level. This method has been applied to a daily generation scheduling problem. The use of an augmented Lagrangian technique, provides satisfactory convergence proper ties to the decomposition algorithm. Moreover, numerical simulations show that compared to a classical deterministic optimization with reserve constraints, this new approach achieves substantial savings.