Sunspot series prediction using adaptively trained Multiscale-Bilinear Recurrent Neural Network

Abstract

A prediction scheme for sunspot series using a Recurrent Neural Network is proposed in this paper. The recurrent neural network adopted in this scheme is the Multiscale-Bilinear Recurrent Neural Network with an adaptive learning algorithm (M-BRNN (AL)). The M-BLRNN(AL) is formulated by a combination of several Bilinear Recurrent Neural Network (BRNN) models in which each model is employed for predicting the signal at a certain level obtained by a wavelet transform. The learning process is further improved by applying an adaptive learning algorithm at each resolution level. In order to evaluate the performance of the proposed M-BRNN(AL)-based predictor, experiments are conducted on the Wolf sunspot series number data and the resulting prediction accuracy is compared with those of conventional MultiLayer Perceptron Type Neural Network (MLPNN)-based and BRNN-based predictors. The results show that the proposed M-BRNN(AL)-based predictor outperforms the MLPNN-based and BRNN-based predictors in terms of the Normalized Mean Squared Error (NMSE).