Research On Long-term Runoff Forecast Of Yalong River Basin Based On Ensemble Learning

The medium and long term runoff forecast is of great significance to the optimal operation of reservoirs and the optimal development and utilization of water resources.Due to the insufficiency and uneven distribution of surface meteorological stations in the Yalong River Basin,it is difficult to obtain accurate rainfall data of the basin surface,and the limited generalization ability of traditional medium and long term runoff prediction models.Therefore,this paper proposes to combine satellite rainfall data with ensemble learning to study the medium and long term runoff prediction.It has certain theoretical significance and practical application value to improve the accuracy of medium and long term runoff forecast.This paper firstly evaluated the accuracy of the Rainfall data of the Tropical Rainfall Measuring Mission(TRMM)3B43 product in the Yalong River Basin based on the measured data of 14 ground stations from 1998 to 2015.Then inverse distance weighted interpolation method is used to interpolate the measured rainfall data in the basin.The interpolation results are compared with the spatial distribution characteristics of satellite rainfall data.On this basis,the catchment area above the dam site of Jinping I is taken as the research area.The correlation analysis between the runoff data and the average satellite rainfall was carried out,and the same procedure was carried out for the measured rainfall at the Jiulong River station.Secondly,historical runoff,monthly mean air temperature,monthly mean water vapor pressure,monthly mean humidity and average satellite rainfall data were used as Forecasting factors.The GBRT(Gradient Boosting Decision Tree)and XGBoost(Extreme Gradient Boosting)based on Boosting algorithm,RF(Random Forest)and ET(Extreme Random Tree)based on Bagging algorithm are respectively adopted.The four algorithms were used as the prediction model to forecast the monthly average inflow flow sequence of Jinping I reservoir,and the prediction results were compared and analyzed.Finally,ET and XGBoost,which have better prediction results,are further selected as primary learners.Then,Lasso regression is used as a secondary learner and 5 fold cross validation is used to model the Stacking ensemble learning.Finally,the model is used to further predict the runoff,so as to get a more accurate forecast result.This provides a new and feasible idea for the middle and long term runoff forecast of Jinping I reservoir.