LEARNING SUNSPOT SERIES DYNAMICS BY RECURRENT NEURAL NETWORKS
Leong Kwan Li · 2003
Sunspot series is a record of the activities of the surface of the sun. It is chaotic and is a well-known challenging task for time series analysis. In this paper, we show that we can approximate the transformed sequence with a discrete-time recurrent neural networks. We apply a new smoothing technique by integrating the original sequence twice with mean correction and also normalize the smoothened sequence to [-1,1]. The smoothened sequence is divided into a few segments and each segment is approximated by a neuron of a discrete-time fully connected neural network. Our approach is based on the universal approximation property of discrete-time recurrent neural network. The relation between the least square error and the network size are discussed. The results are compared with the linear time series models.