A hybrid neural architecture and its application to temperature prediction

Abstract

We first investigate the effectiveness of multilayer perceptron networks for prediction of atmospheric temperature. To capture the seasonality of atmospheric data we then propose a hybrid network, SOFM-MLP, that combines a self-organizing feature map (SOFM) and multilayer perceptron networks (MLPs). The architecture is quite general in nature and can be applied in other application areas. We also demonstrate that use of appropriate features can not only reduce the number of features but also can improve the prediction accuracies. © Springer-Verlag Berlin Heidelberg 2003.