DE TRAVAIL 2002-013 A NEURO-DYNAMIC PROGRAMMING APPROACH FOR STOCHASTIC RESERVOIR MANAGEMENT

We propose an approach based on neural networks for optimizating a single hydroelectric reservoir. A stochastic neuro-dynamic programming algorithm is used to approximate the future value function by a neural function. The latter is used in deriving the optimal policy. The approximation architecture, based on the feedforward network, gives very smooth approximate functions even with a coarse discretization of the state and action variables. The hydroelectric reservoir model presented in our study assumes a piecewise linear reward of the electricity produced and takes into account the turbine head effects and the stochastic inflows. The method is illustrated with a numerical example.