Environmental Time Series Prediction by Improved Classical Feed-Forward Neural Networks

Abstract

The water quality at the issue of a wastewater treatment plant (WWTP) is a complex work because of its complexity and variability when conditions suddenly change. Two main techniques has been used to improve classical feed-forward neural network. First, a classical adaptative gradient learning rule has been complemented with a Kalman learning rule which is especially effective for noisy behavioral problems. Second, two independent variable selection components -based on genetic algorithms and fuzzy ranking- have been implemented to try to improve performance and generalization. The global study shows that reliable results are obtained which permit to guarantee that neural networks are a confidence tool on this subject.