Evaluation of Neural Network Streamflow Forecasting on 47 Watersheds

This study is designed to compare 1day ahead streamflow forecasting performance of multiple-layer perceptron (MLP) networks implemented at a daily time step for 47 watersheds spread across France and Central United States. In order to keep the task to manageable proportions, a large sample of test watersheds asks for a reduction of the number of steps in the network implementation procedure. This is achieved by eliminating the long trial and error process of input selection. Results show that it is feasible to obtain good 1day ahead streamflow forecasting performance from simple MLPs and input vectors consisting solely of the last observed streamflow and a predetermined range of precipitation observations that is roughly equal to the time of concentration of the watersheds. Also, intuitive preprocessing such as differencing the streamflow noticeably improves the forecasting performance in almost all instances. On the other hand, consideration of the potential evapotranspiration as an additional input decreases the MLP’s performance in the majority of instances. Finally, it is noteworthy that there is a general trend between the watershed runoff coefficients and the ability of the MLPs to correctly map 1day ahead streamflows.