Long-term stream-flow prediction and water risk identification by the data-driven support vector regression with a dynamic stream deficit index

ABSTRACT: In this paper, a data-driven Support Vector Regression with a Dynamic Stream Deficit Index (SVR-DSDI) model is proposed to identify the water risk event in a long-term lead time for reservoir regions. In the SVR-DSDI model, the Support Vector Regression (SVR) is used to predict the stream-flow and the precipitation in the subsequent days, whereas Dynamic Stream Deficit Index (DSDI) is calculated as the drought or flood degree of the prediction risk. In order to verify the accuracy and the reliability of our proposed scheme, a conventional Back Propagation Neural Network (BPNN) model is compared with the SVR-DSDI in monthly stream-flow forecasting. The experimental results reveal that the predictabilities of the SVR-DSDI with adaptive parameter are significantly superior to the BPNN models. Furthermore, the DSDI as the water risk degree is consistently reliable as an indicator of the true risk.