Study On Mid-long Term Runoff Forecast In Weihe River Basin

Due to the influence of a series of factors such as climate,underlying surface and human activities,runoff sequence is often characterized by randomness,skewness and nonlinearity.Timely and accurate runoff prediction is of great practical significance and application value for water resources allocation management and reservoir operation decision-making.Weihe river is the largest tributary of the Yellow River.Contributing 18%of the water resources in Shaanxi province,Weihe river basin supports 56% of the province's farmland,72% of the irrigated area and 75% of the gross national product.Medium-long term runoff forecast will provide scientific decision-making basis for the development and utilization of water resources,flood control,drought resistance and power generation in the Weihe river basin.In this paper,considering the feature variation of runoff sequence,the monthly runoff data of six controlled hydrological stations in Weihe river basin,including Linjiacun,Weijiabao,Xianyang,Huaxian,Zhangjiashan and Zhuangtou,were selected.Three runoff sequence treatment methods including the Box-Cox transformation(BC),the Min-Max standardization(MM)and the wavelet analysis(WD),two prediction factor selection methods including the Gray correlation method and Lasso regression method,and three models including the BP neural network,the projection pursuit regression and the support vector regression model were combined.24 monthly runoff forecast model were established to study the middle-long-term runoff forecast.Meanwhile,all models were comprehensively evaluated and selected according to three prediction error index and the main results are as follows:(1)The prediction factors selection methods of the BC-Lasso,the BC-Gray,the MM-Lasso and the MM-Gray were constructed.For the same runoff prediction model,it can be seen from the evaluation indexes of the six hydrological stations in the verification period that the ranking of the comprehensive prediction effect of the four methods from the best to the worst is the BC-Lasso,the BC-Gray,the MM-Lasso,the MM-Gray,respectively.The results show that the method of normalizing the data with the Box-Cox transformation and using the Lasso regression to optimize the prediction factor set can effectively improve the prediction effect.(2)Prediction factors selection methods of the WD-BC-Lasso,the WD-BC-Gray,theWD-MM-Lasso and the WD-MM-Gray were constructed based on wavelet analysis.The ranking of the comprehensive forecasting effect from the best to the worst is the WD-BC-Lasso,the WD-BC-Gray,the WD-MM-Lasso,the WD-MM-Gray,respectively.The selection method based on the wavelet analysis preprocessing technology is better than that without preprocessing technology.The research results show that the data decomposition and reconstruction based on the wavelet analysis has improved the prediction effect of the model.(3)The BP neural network model,the projection tracing regression model and the support vector regression model were constructed.For the same runoff sequence treatment and prediction factor selection method,it can be seen from the evaluation indexes of the six hydrological stations in the verification period that the comprehensive prediction effect of the three models was ranked from the best to the worst as SVR,BP and PPR.The research results show that the support vector regression model can effectively realize the global optimal solution under finite samples and has a good generalization ability.For the 24 prediction methods of 6 hydrology stations,although the models and quantities that meet the prediction requirements are different,the WD-BC-LSVR model shows good prediction accuracy and stability in monthly runoff prediction of Weihe river basin by comprehensive comparison.During the verification period,the MRE of the six hydrothermal stations was less than 17%,and R was bigger than 0.97,and Ens was greater than 0.93,indicating that the simulation effect of the WD-BC-LSVR model was better than that of other models and had obvious advantages.Meanwhile,The optimal model of 5stations in Linjiacun,Weijiabao,Xianyang,Huaxian and Zhangjiashan is the WD-BC-LSVR model.Although the model is not optimized in Zhuangtou station,its prediction effect can still meet the requirements.To sum up,the WD-BC-LSVR is the optimal model among 24 forecasting methods,which can be used for monthly runoff prediction of Weihe river basin.