Multiscale RBF Neural Network for Forecasting of Monthly Hake Catches off Southern Chile

Abstract

We present a forecasting strategy based on stationary wavelet transform combined with radial basis function (RBF) neural network to improve the accuracy of 3-month-ahead hake catches forecasting of the fisheries industry in the central southern Chile. The general idea of the proposed forecasting model is to decompose the raw data set into an annual cycle component and an inter-annual component by using 3-levels stationary wavelet decomposition. The components are independently predicted using an autoregressive RBF neural network model. The RBF neural network model is composed of linear and nonlinear weights, which are estimates using the separable nonlinear least squares method. Consequently, the proposed forecaster is the co-addition of two predicted components. We demonstrate the utility of the proposed model on hake catches data set for monthly periods from 1963 to 2008. Experimental results on hake catches data show that the autoregressive RBF neural network model is effective for 3-month-ahead forecasting. Index Terms—Neural network, forecasting, nonlinear least squares.