Real-time short-term natural water inflows forecasting using recurrent neural networks

Accurate, time and site-specific forecasts of natural inflows into hydropower reservoirs are highly important for operating and scheduling. This paper investigates the effectiveness of recurrent neural networks (RNN) for real-time short-term natural water inflows forecasting. The models use antecedent inflows and precipitation data, and actual weather descriptors to generate short-term (1-7 days ahead) natural inflow forecasts for a specific hydroelectric reservoir. The input variables are exactly the same as those previously used for an autoregressive moving average model with exogenous variables (ARMAX) and for a conceptual model (PREVIS). The RNN are trained using the early stopped training technique with the Levenberg-Marquardt backpropagation. The experimental results show that the performance of RNN using the early stopped training approach outperforms the traditional stochastic model and the available conceptual model. Particularly, the RNN have shown better forecasting capabilities for the last 3 of the seven days ahead forecasts.