Learning disaggregation technique for the operation of long‐term hydroelectric power systems

This paper describes a nonlinear disaggregation technique for the operation of multireservoir systems. The disaggregation is done by training a neural network to give, for an aggregated storage level, the storage level of each reservoir of the system. The training set is obtained by solving the deterministic operating problem of a large number of equally likely flow sequences. The training is achieved using the back propagation method, and the minimization of the quadratic error is computed by a variable step gradient method. The aggregated storage level can be determined by stochastic dynamic programming in which all hydroelectric installations are aggregated to form one equivalent reservoir. The results of applying the learning disaggregation technique to Quebec's La Grande river are reported, and a comparison with the principal component analysis disaggregation technique is given.