Rainfall-runoff modeling using Artificial Neural Network technique for Baitarni river in Orissa – A Research Report by National Institute of Hydrology

The research study attempts to develop a rainfall-runoff model using the Artificial Neural Network (ANN) technique for the Baitarni river in Orissa. A detailed review of the research work in the area of interest revealed that the approach of neural computations was very effective in developing the required model, due to its various advantages. Accordingly, three candidate models based on ANN architecture were developed for the study area, to represent rainfall-runoff transformation.

The architecture of all these models was determined based on a trial and error procedure and by examining the goodness of fit statistics. An auto-correlation and partial auto-correlation analysis of the standardized daily flow series suggested that the flow at time ‘t’ was highly correlated to the previous three days flows viz (Q_t-1, Q_t-2, Q_t-3). These parameters were included in the input vector of the network, apart from 5-days rainfall series prior to the day, at which the flow was to be predicted. The number of rainfall patterns in the input vector was finalized by trial and error procedure.

Statistical analysis was done on the performance of each model in estimating the runoff. The study revealed that an ANN with a radial basis function algorithm was able to model the Rainfall-Runoff transformation more accurately than a back propagation network. However, for estimation of peak flows a BPN with twelve neurons in the hidden layer was found efficient. The results from the Radial Basis Function model were compared with the results of existing models such as Sacramento model and regression equations developed, and the performance of the former was found to be superior to others.

Hence, it was concluded that the Radial Basis Function network model developed for rainfall-runoff process in the Baitarni river basin might be employed for water resources planning. While such a model is not intended as a substitute for physically based models, it can provide a viable alternative when the hydrologic application requires that an accurate forecast of stream flow be provided using only the available input and output time series data, and with relatively little conceptual understanding of the hydrologic dynamics of the particular basin under investigation.

Download the report here: