Prediction of Daily Streamflow Using Various Kernel Function Based Regression: A Case Study in India
Abstract River daily discharge estimation and modeling considers an important step for scheduling and planning different water resources for sustainable socio-economic development. In the current work, four techniques of Gaussian processes regression (GPR): Polynomial Kernel, Radial Basis Function Kernel, Normalized Polynomial Kernel, and PUK Kernel, were used to model the daily discharge. Hydrological-datasets containing daily-stage (m) and discharge (m3/sec) were gathered over the period from 2004-2013. The datasets were divided into two sections: (i) models training containing 70% (2004-2010) of the total data and (ii) remaining 30% (2011- 2013) were for testing. Comparing all the four developed models, our findings show that the superlative model was the PUK-Kernel model with a correlation coefficient (r) of 0.96, MAE of 36.70 m3/s, RMSE of 90.92 m3/s, RAE of 17.50 %, RRSE of 26.05 % in the training period. Whereas, it performed equally well in the testing period with r = 0.97, MAE = 44.84 m3/s, RMSE = 95.05 m3/s, RAE = 17.98 %, RRSE = 24.94 % in the testing period. Our findings can be included that GPR-PUK was more accurate and stable than other models, and can be used to help water-users, decision-makers, development-planners for managing water resources and achieving sustainable development.
