Impact Assessment of Antecedent Hydro-meteorological Parameters Data on the Performance of Support Vector Machines (Regression) Based Stream Flow Prediction Model Integrating Genetic Algorithm

Abstract

Accurate and reliable river flow predictions are obvious for appropriate planning, development and management of water resources, particularly for a country like Pakistan where cultivation is mostly by canal irrigation system. This is particularly important for sustainable socio-economic growth, proper management of the canal system and flood mitigation under changing climatic conditions. In this study, thirty years’ (1985 – 2014) monthly temperature, precipitation and streamflow data from Astore sub-basin of the Upper Indus River Basin, UIRB in Pakistan have been analysed. The streamflow of the Astore River, which is a tributary of the Indus River, is predicted ahead of time, considering the impact of antecedent precipitation, the temperature and streamflow data. During the recent past decades, artificial intelligence-based modeling with several categories of models has been presented as an important technique for the prediction of hydrological phenomenon. In this paper, the performance of four Support Vector Machines Regression (SVR) models have been probed to predict the streamflow of Astore River. The Four SVR model types were compared on the basis of radial basis function, polynomial, linear and sigmoid kernels. Number of input combinations with input variables (temperature, precipitation, and streamflow) with reference to time lag were determined by Genetic Algorithm test. The best input combination for SVR models was identified using a genetic algorithm upon the bases of the smallest values of gamma and Standard Error. The Nash-Sutcliffe efficiency and Mean Bias error were used to evaluate the performance of SVR Models. The SVR model, based on radial basis function kernel forecasted the stream flows with higher accuracy as compared to the other kernels.