Developing Methods to Train Neural Networks for Time-Series Prediction in Environmental Systems

This paper proposes the local interaction method to train neural networks for predicting future variable values of environmental system. Time-series data including soil, stream water and climatic variables were measured hourly over half of a year at two observation spots in Qingpu district, 45 kilometers west to Shanghai city. Three different methods, including our biologically plausible method, have used the data sets to train neural networks. The temporal pattern recognition capabilities for these methods were compared. Our method was proved more competitive than the other two traditional methods in using large data sets to detect patterns and predict events for complex environmental systems.