Influence of spatial variation in precipitation on artificial neural network rainfall-runoff model

Modeling rainfall-runoff processes is a very challenging task due to data collection, time, money, and technology constraints. Artificial neural networks (ANNs) are modeling tools that can quickly adapt and learn input-output relationships for many different engineering problems. An Elman-type recurrent ANN was trained to simulate observed streamflow for Fountain Creek at Pueblo, CO, using varying amounts of spatial precipitation information. Nine zones were originally delineated within the watershed draining to Fountain Creek at Pueblo based on estimated overland flow travel time. Five different spatially varying scenarios were modeled: scenarios containing 9 zones, 6 zones, 3 zones, 2 zones, and 1 zone. Each scenario was trained and simulated 100 times, each with randomly generated initial weights. Spatial variability in precipitation data allows the ANN to achieve better performance when simulating the training dataset. However, when applied to the validation and testing time periods, ANN performance generally decreases with additional spatially variability. In addition to exploring results of the ANN rainfall-runoff model, the application of geographical information systems to rainfall-runoff input processing is demonstrated.