Artificial Neural Networks for Forecasting Watershed Runoff and Stream Flows

Abstract

This research demonstrates an application of artificial neural networks (ANN) for watershed-runoff and stream-flow forecasts. A watershed runoff prediction model was developed to predict stormwater runoff at a gauged location near the watershed outlet. Another stream flow forecasting model was formulated to forecast river flows at downstream locations along the same channel. Input data for both models include the current and preceding records of rainfall and stream flow gathered at the watershed outlet and downstream locations. Computational algorithms for both models were based on a commercially available software. A case study was conducted on a small urban watershed in Greensboro, North Carolina. These two ANN-hydrologic forecasting models were successfully applied to provide near-real-time- and near-term-flow predictions with lead times starting from the present time and advancing to a few hours later on 15-min increments. An important aspect of this research has been the development of methodology for input data organization, model performance evaluation, and ANN processing techniques. Encouraging results obtained indicate that ANN-hydrologic forecasting models can be considered an alternate and practical tool for stream-flow forecast, which is particularly useful for assisting small urban watersheds to issue timely and early flood warnings.