Performance comparison of feedforward neural networks applied to stream flow series forecasting

Abstract

Feedforward neural networks are those in which the input signal follows only one direction: from the input layer to the output layer, passing through all the hidden layers, in contrast with recurrent architectures. The main examples of this class are the Multilayer Perceptron (MLP) and the Radial Basis Function network (RBF). Recently, other model of this type has received significant attention: Extreme Learning Machines (ELMs). Nonlinear mapping problems, like time series forecasting, can be adequately solved by these methods. In this work, the aforementioned architectures are employed in the prediction of monthly seasonal streamflow series of important Brazilian hydroelectric plants, for different forecasting horizons. The results showed that all the proposals are efficient in solving the task, but, interestingly, the RBF achieved the best performance in most cases. However, the computational cost associated with the training process of the ELM is much smaller than the others.