Neuroevolution Applied to River Level Forecasting Under Winter Flood and Drought Conditions

Robert J. Abrahart, Linda M. See, Alison J. Heppenstall

School of Geography, University of Nottingham, Nottingham, NG7 2RD, UK

Abstract

JavaSANE is used to develop a series of neural network river forecasting solutions based on the principles of cooperative coevolution. Neural network rainfall-runoff models are constructed for a downstream gauging station on the River Ouse in northern England. Forecasting solutions are developed for different hour lead times: T+6, T+12, T+18, and T+24. The solutions are first developed on a sum squared error objective function that is thereafter modified on a T+6 model to calculate sum squared error computed on the bottom 50% and top 50% of the dataset. The intention is to determine whether improvements in lower or higher magnitude forecasts can be achieved and to assess the effect that such modifications would have on the remainder. Each model was trained on a hydrometeorological dataset for one winter period and tested on two independent winter period datasets: one test set represented a normal winter period with major floods; the other test set contained winter drought conditions. The potential benefits of dedicated targeting is demonstrated with improved results being obtained for lower magnitude events, under winter drought conditions, with limited detrimental side effects occurring.

Keywords: evolutionary algorithms, neural networks, rainfall-runoff modeling