An assessment of multi-layer perceptron networks for streamflow forecasting in large-scale interconnected hydrosystems

This work analyzes the use of artificial neural networks in the short-term streamflow forecasting for large interconnected hydropower systems. The state-of-the-art optimization algorithms, activation functions, and weight initialization techniques are investigated together with classic methods. We present an algorithm to define the neural network inputs in large hydrosystems and apply it to create models for 55 major hydro plants located in the Parana Basin, which contribute to more than 30% of the total power generated in Brazil. The paper also compares the performance of the neural networks with the hydrological models that are currently used by the independent system operator to define the dispatch of the electric power generators. Our results show that, overall, the neural network models provide more accurate forecasts than the hydrological models used by the Brazilian System Operator. Finally, the paper discusses the contributions of historical rainfall information in the forecasting of streamflow while using neural network models.