Multiple Input Single Target Streamflow Forecast by Neurowavelet Networks

Abstract

In this work a methodology of time series analysis and forecasting is implemented and tested against streamflow data measured from the Mississippi River. Measurements from a set of twelve data stations along the river is used in various configuration to test the potential improvement in predictive efficiency from incorporation of a network of sensors. The forecast system is based on combining the discrete wavelet transform with the well-established feed forward non-linear auto regressive design multilayer perceptron artificial neural network. These networks, referred to as neurowavelets, are able to improve predictive ability. The novel neurowavelet system is capable of processing multiple external inputs with separate selection of wavelet family and resolution. Improved performance and stability is gained through an evolutionary genetic algorithm applied to the backpropagation training process combined with the use of ensemble learning. The system is tested both with and without performing multiresolution analysis of the inputs and a clear advantage is seen when introducing the wavelet decomposition. Results show improvement from the refined model and best forecast performance is obtained from use of the full sensor cascade and the Coiflet wavelet family.