Prediction of Urban Stormwater Runoff in Chesapeake Bay Using Neural Networks

Abstract

Runoff carries pollutants such as oil, heavy metals, bacteria, sediment, pesticides and fertilizers into streams or groundwater. The combined impacts of hydrologic changes and water pollution can be disastrous for streams and rivers in urban areas and the Chesapeake Bay. Therefore, evaluations of stormwater runoff are imperative to enhance the performance of an assessment operation and develop better water resources management and plan. In order to accomplish the goal, a recurrent neural network based predictive model trained by the Levenberg-Marquardt backpropagation training algorithm is developed to forecast the runoff discharge using the gage height and the previous runoff discharge. The experimental results showed that Levenberg-Marquardt backpropagation training algorithm proved to be successful in training the recurrent neural network for the stormwater runoff prediction. Based on the comparison studies about the impact of discharge and gage height on the runoff forecast accuracy, it was found that when both the previous discharge and gage height were used, the network achieved lower mean squared error, and better time series response than the case when the gage height is the only input or target.