Comparison of ANN with other methods for rainfall runoff modeling of a catchment with limited data

In this study, two black-box rainfall-runoff models namely, the Linear Perturbation Model and an Artificial Neural Network and a conceptual catchment model were applied to determine their suitability to model rainfall-runoff process in a catchment having limited data. In case of the linear perturbation model, the memory length was selected as the minimum value beyond which there is no appreciable improvement in model performance. A three layer feed forward artificial neural network was trained using the back propagation algorithm. The conceptual model used here is based on linear storages and its parameters were determined using a combination of automatic and manual calibration. Performances of these three models were evaluated through several statistical criteria. Results show that the artificial neural network model performed significantly better than the conceptual and linear perturbation models in the situation where data are scanty and of poor quality. Further, the performance of the artificial neural network improved when the data of individual rainfall stations was used as input rather than the mean annual rainfall for the catchment.