Runoff Prediction of Bharathapuzha River Basin using Artificial Neural Network and SWAT Model

Abstract

An attempt was made to model the non-linear system of rainfall-runoff process from Bharathapuzha River basin using an information processing paradigm, Artificial Neural Network (ANN). The results were compared with the outputs of the semi-distributed, physically-based SWAT (Soil and Water Assessment Tool) model. The ANN modelling was done using back propagation learning algorithm, tan sigmoid transfer function, and model input strategy having rainfall and other climatic variables as input by assigning number of layers as 5, 10, 15, 20, 25, 30, and 40. Different models were evaluated with respect to coefficient of correlation (r), coefficient of determination (R2 ), and root mean square error (RMSE). Among the ANN models, ANN-BP-A-5 (six input variables, 5 hidden layers) performed best, followed by ANN-BP-A40 (six input variables, 40 hidden layers). Comparison of ANN predicted runoff of the best models (ANN-BP-A-5 and ANNBP-A40) with the SWAT predicted runoff revealed that the simulated runoff using SWAT was more correlated to observed runoff than ANN predicted runoff. The ANN models underestimated the flow during the rainy season, and gave an overestimation during the summer season. However, the R2 values of 0.666 and 0.649 obtained for ANN-BP-A-5 and ANN-BP-A40, respectively, indicated that the performances of ANN models were satisfactory and ANN model can also be used for runoff prediction in data scarce areas.