Adaptive architecture of polynomial artificial neural network to forecast nonlinear time series

Abstract

There are two important ways in which artificial neural networks are applied for dynamic system identification: preprocessing the training values, and adapting the architecture of the network. The article describes an adaptive process of the architecture of Polynomial Artificial Neural Network (PANN) using a genetic algorithm (GA) to improve the learning process. The optimal structure is obtained without previous knowledge of the behavior of the system to be identified. Due to the nature of the structure of PANN, it is possible to extract the necessary information of the nonlinear time series in order to minimize the training error. The importance of this work lies on adapting the architecture of PANN and processing the necessary inputs to minimize this error at the same time. The training error is compared with other networks used in the field to forecast chaotic time series.