A rain-streamflow model for prediction of limnimetric behavior of reservoirs using artificial neural networks

摘要

This paper proposes a rain-streamflow model for twenty-four hours forecast of hydroelectric reservoir level using Artificial Neural Networks. The model was developed using Matlab? Neural Network Toolbox software. A multi-layer perceptron (MLP) neural network, trained with the Bayesian regularization algorithms, was chosen. The inflow prediction task is important for reliable and economical operations of electricity generation projects as well as any project that requires the representation of hydrological processes. Many conventional methods have been used to perform water inflow forecasting, however, such methods require expert intervention for adjusting physical model parameters. Artificial Neural Networks have the advantage of automatically adjusting their parameters, better correcting the prediction errors and with less computational time. The developed model used the historical series of the Preto river together with a network of hydrometric stations for training and verification. For the analyzed case the developed model obtained satisfactory results, reaching mean percentage errors of the order of 0.02% for the forecast horizon of 24 hours. The use of precipitation brought gains to the model, since the forecasts without this information presented errors 31% higher.