Prediction of hourly runoff in the Savitri River basin in India: Use of a local approximation approach

In this study, the concept of nonlinear dynamics is used to predict runoff in the Savitri River basin, Maharashtra, India. Hourly runoff from four stations in the basin, namely Kangule, Bhave, Birwadi, and Kokkare, are studied. The nonlinear prediction method with a local approximation approach is employed, and one-hour-ahead runoff predictions are made. The method uses (1) reconstruction of the single-dimensional runoff series in a multi-dimensional phase space; (2) determination of the Euclidean distances between the reconstructed vectors; and (3) prediction using a nearest-neighbor local approximation approach, with consideration of different number of neighbours. For each of the four streamflow stations, data observed during the period 2000–2009 are used for phase space reconstruction, and predictions are made for the year 2010. Three statistical evaluation measures, correlation coefficient (CC), Nash-Sutcliffe efficiency (NSE), and normalized root mean square error (NRMSE), are used to determine the performance of the method. The prediction results for the four stations indicate very good accuracy, with the CC values ranging between 0.980 and 998, the NSE values between 0.961 and 0.995, and the NRMSE values between 0.010 to 0.014. The optimal embedding dimensions (i.e. the dimensions yielding the best predictions) for the Kangule, Bhave, Birwadi, and Kokkare are 9, 13, 6, and 9, respectively. These dimensions suggest that the complexity of the dynamics of hourly runoff in the Savitri River basin is medium-to-high dimensional. The outcomes from the present study are certainly encouraging to further enhance the application of nonlinear dynamic concepts for studying the runoff dynamics in the Savitri River basin.

Keywords: Runoff prediction, Nonlinear dynamics, Chaos, Phase space reconstruction, Local approximation prediction, Savitri River basin