École Nationale Supérieure de
Techniques Avancées

Abstract

Hydroelectricity plays an important role in the electricity demand supply balance. However, it is a rather random and limited ressource. At EDF, mid-term (that is to say annual basis) optimisation of hydropower reservoirs is calculated with a stochastic dynamic programming algorithm. This algorithm hasshowntobeeffectiveinmanagingvalleyswithfewreservoirsbutrequieressimplifyingassumptions to deal with larger valleys. Indeed, its complexity grows exponentially with the number of optimised storages. Directpolicysearchisanalternativemethodofstochasticdynamicprogramming.Apolicyisafunction which determines actions depending on state variables of the optimized system. Direct policy search aims at choosing a set of parametrized functions and searching the best policy among this set. Optimisation of large size valleys is made possible through this new approach. This report is devoted to direct policy search study. Specifically, a Primal-Dual algorithm will be implemented to optimise storagesunderconstraints.Theresultswillbecomparedtostochasticdynamicprogrammingonesinorder to measure potential stakes of the direct policy search applied to hydropower systems.