On Wavelet Neural Networks and River Flow Forecasting

Abstract

A method of river flow modeling and forecast is implemented, and results are presented to provide comparisons based on different techniques and training parameters. Here we implement a forecast based on the well-established feed forward back propagation design multilayer perceptron artificial neural network. In order to improve predictive ability, two new methods are designed to incorporate the multi-resolution information from a Daubechies type wavelet transform as input to the network. The novel methods are compared with the existing one in a case study to assess the performance of the wavelet neural networks, and to obtain results to help guide future network design and select of training parameters. The new predictive network design is inspired by existing methods but adds more repeatability and stability to the result. By using a genetic algorithm for selecting trained networks and averaging the results of many trials, we can incorporate the inherent randomness created from network training. In this case study, we combine wavelet analysis and artificial neural networks to perform river flow forecasting of the Tittabawassee River. Our results are superior to some existing methods.