The study evaluates the applicability of Artificial Neural Networks (ANN) in rainfall-runoff process modeling for the Ajay river basin, Jharkhand to predict the 6-hour ahead runoff at Sarath. Initially three flood events were considered for training, to find out the weights between different layers of the network. The back propagation algorithm has been used for optimization of weights. The developed ANN was validated for rest of the flood events.
The validation results suggested the updating of the existing weights, which may represent the rainfall-runoff process in the catchment. More flood events were considered for training the network, to determine weights that may represent the rainfall-runoff processes in the catchment more realistically. Six sets of flood events were used in turn for training the network. The trained networks were then validated based on the remaining flood events.
The results showed that the weights determined by considering flood events FH1, FH2, FH3, FH4, FH6 (case 2) in training result in least r.m.s.e in validation. A check has been performed to observe any significant change in the weight matrices of case 2 by including one more flood event (FH5) in the training process. No significant change has been observed in the weight matrices of case 2 for the available set of data. Thus the weights determined in case 2 can be used for issuing the 6-hour ahead forecast at Sarath gauging site of Ajay river basin for the given condition.
However, as the rainfall-runoff process is not stationery and every flood event has its own characteristics there is always a need for updating the existing model, whenever new flood events become available. The ANN model can perform better if more flood events with significantly low noise component are made available for modeling the rainfall-runoff process.
Download the report here: