Neural network-based long-term hydropower forecasting system

Artificial neural networks are alternatives to stochastic models even if th e optimization of their architectures remains a tricky problem. Two differe nt approaches in long-term forecasting of potential energy inflows using a feedforward neural network (FNN) and a recurrent neural network (RNN) are p roposed The problem of overfitting, particularly critical for limited hydro logic data records, is addressed using a new approach entitled optimal weig ht estimate procedure (OWEP). The efficiency of the two models using OWEP i s assessed through multistep forecasts. The experiment results show that in general, OWEP improves the models' performance and significantly reduces t he training time on the order of 60 percent. The RNN outperforms the FNN bu t costs about a factor of 2 longer in training time. Furthermore, the neura l network-based models provide more accurate forecasts than traditional sto chastic models. Overall, the RNN appears to be the best suited for potentia l energy inflows forecasting and therefore for hydropower systems managemen t and planning.