Neural networks and seasonal time-series prediction

Seasonal time series with dominant frequency components are very common in various information and control systems, biomedicine and environmental engineering. Methods for their analysis, modelling and prediction include both linear and nonlinear structures using autoregressive models, adaptive systems and neural networks. The paper presents basic algorithms for preprocessing of such signals including their filtering and subsampling at first to enable the following efficient signal approximation, modelling and prediction. The main part of the paper presents the comparison of linear autoregressive models and neural networks with different transfer functions including wavelets as well. Resulting methods presented in the paper are applied for processing of real values representing the river Elbe flow at a selected measuring station to study their modelling and prediction. The paper compares results obtained both by classical and adaptive methods using system history and appropriate learning methods for model optimization.