Comparative Evaluation of Neural Network and Regression Based Models to Simulate Runoff and Sediment Yield in an Outer Himalayan Watershed

The complexity of rainfall-runoff-sediment yield hydrological processes remains a challenge for runoff and sediment yield prediction for large mountainous watersheds. In this study, a simple Non-Linear Dynamic (NLD) model has been employed for predicting daily runoff and sediment yield by considering the watershed memory based rainfall and runoff, and rainfall-runoff and sediment yield, respectively. The results were compared with two commonly used Artificial Neural Network (ANN) and Wavelet based ANN (WNN) models by taking maximum input parameters of values of time memory for rainfall, runoff, and sediment yield derived from the developed NLD model through step-wise regression. The feed forward ANN models with back propagation algorithm was used. Twenty-six years’ daily rainfall, runoff, and sediment yield data of Bino Watershed, Uttarakhand, were used in this study. The coefficient of determination, root mean square error, and model efficiency were adopted to evaluate the model’s performance. The results revealed a better performance by the ANN and WNN rainfall-runoff models compared to the NLD, however, NLD rainfall-runoff-sediment model showed higher efficiency than the ANN and WNN models in case of considering whole time series data. Under-prediction of sediment yield by all the models resulted from sudden landslides/flash floods in Himalayan Watersheds. The study showed that though WNN was better than ANN and NLD, its application cannot be generalized for entire mountainous watersheds. Again, criteria for successful selection of a useful sub-component in WNN need to be developed. The study also indicates the greater capturing power of WNN for simulation of extreme flows with lowest percent-error-peak-flow values.

Keywords:

Dynamic , Mountainous watershed , Neural Networks , Peak flow , Time lag