Application of a recurrent neural network to rainfall-runoff modeling

The lumped daily rainfall-runoff process for the Leaf River Basin in Mississippi was modeled using two different Artificial Neural Network (ANN) model structures. Our results indicate that both structures, the popular Three Layer Feedforward Neural Network (TLFNN) and the Recurrent Neural Network (RNN), perform well. However, the TLFNN requires trial-and-error testing to identify the appropriate number of time-delayed input variables to the model. Further, it is not suitable for distributed watershed modeling; i.e., when distributed precipitation information (multiple gages or radar images) is available. The RNN structure provides a representation of the dynamic internal feedbacks loops in the system, thereby eliminating the need for lagged inputs and resulting in a reduction in the number of network weights (and hence training time). The suitability of RNN's for distributed watershed modeling is discussed.