Mid-long Term Runoff Prediction Based on a Lasso and SVR Hybrid Method

The longer the forecast lead time,the better the effect of decision-making support for reservoir operation schedule.The main challengings of mid-long term runoff forecast are randomness of runoff time series and lack of reliable weather forecasting with same forecast lead time.Consequently,many data-driven models are proposed to simulate monthly runoff time series,and the inputs of those model are mostly previous runoff series and climatologic factors.However,the influence of input variables selection on simulation results are seldom discussed,although input variables selection is much crucial.In this paper,the lasso method is used to facilitate predictors selection for developing support vector regression (SVR) model for monthly runoff prediction at the Longyangxia reservoir,which is called LSVR model.The application results of the developed LSVR model are compared with those of the traditional SVR model to investigate influence of predictor selection on forecasting results and the performance of LSVR model.The results demonstrate that runoff predictors selection has significant impact on forecasting effectiveness of both two models.The results indicate that the performance of the developed LSVR model is significant better than that of the traditional SVR model in both validation and verification periods,since the LSVR model can preferentially choose positive predictors and weaken the influence of negative predictors.The simulation accuracy of the runoff forecast with one-month lead time from January 2010 to October 2016 by the LSVR model has been improved by 13.09% with the mean squared error (MSE) evaluating indicator,comparing with that by of the SVR model. © 2018, The Editorial Board of Journal of Basic Science and Engineering. All right reserved.