Study On Medium-and Long-term Runoff Forecast In Songhua River Basin

The selection of potential predictors is one of the key issues in applying data-driven runoff forecasting models to actual projects.Researchers at home and abroad mostly focus on the optimization and improvement of forecast models,the study of model parameter optimization methods,and the comparative analysis of different forecast models and different parameter optimization methods.However,there are few studies on the comprehensive comparative analysis and applicability of factor optimization schemes..In order to extract more effective forecast information and improve the simulation accuracy of the model,this paper introduces 74 atmospheric circulation indexes,and participates in the optimization of runoff forecast input factors together with the precipitation and runoff hydrological sequences of representative stations in the basin.Correlation analysis method(CA),Principal component analysis based on correlation analysis(CA-PCA),mutual information method(MI),and principal component analysis based on mutual information(MI-PCA).Different screening results under 4 factor optimization schemes are used as multiple linear regression(MLR),support vector regression model(SVR),BP neural network model(BP)input,predict the monthly average flow of the Jiamusi Hydrological Station in the Songhua River Basin,and obtain the screening plan and the potential influencing factors suitable for the Songhua River Basin runoff forecast Forecast model.The main conclusions are as follows:(1)In the Songhua River Basin,there are obvious spatial characteristics between runoff and atmospheric circulation indicators,and the correlation gradually increases from upstream to downstream.The atmospheric circulation index when the predetention period is one month,the monthly average runoff of the three upstream stations(Fuyu,Dalai and Harbin)and the monthly average rainfall of Jiamusi Station have the greatest correlation with the monthly average runoff of Jiamusi,which confirms that the atmosphere The lag period before the best forecast of circulation index and hydrological influence factor is one month.(2)Under the MLR model,the forecast effect is MLR-CA-PCA>MLR-CA>MLR-MI-PCA>MLR-MI.In the CA scheme,the forecast effect is MLR-CA7>MLR-CA4>MLR-CA3,but the pass rate does not reach the C level,and the overall effect is not ideal;in the CA-PCA scheme,MLR-CA-PCA4 has the best fitting effect,MLR-CA-PCA3 and MLR-CA-PCA7 fitting curves basically overlap,the forecast results of the three sets of plans have reached the level of C-level forecast;MI and MI-PCA have poor results.Under the SVR model,the forecast effect is SVR-CA-PCA>SVR-C A>SVR-MI-PCA>SVR-MI.In the CA scheme,SVR-CA7>SVR-CA3>SVR-CA4,in w hich the QR of SVR-CA7 is 71.67%,reaching the second-class level;in the CA-PC A scheme,SVR-CA-PCA7>SVR-CA-PCA4>SVR-CA-PCA3,in which the QR of SVR-CA-PCA7 is 73.33%,reaching the second-class level,which is significantly bett er than the other two programs;MI and MI-PCA are also less effective.Under the BP neural network model,the prediction effect is BP-CA-PCA>BP-CA>BP-MI-PCA a nd BP-MI.In the CA scheme,BP-CA3>BP-CA7>BP-CA4;in the CA-PCA scheme,BP-CA-PCA7>BP-CA-PCA4>BP-CA-PCA3,where the RMSE of BP-CA-PCA7 is 827.58 m~3/s,R~2 is 0.81,QR is 78.33%.MI and MI-PCA are equally inferior.(3)The CA optimization factor scheme has achieved the best prediction effect under the fitting of the SVR model,followed by the BP neural network model,and finally the MLR model.The CA-PCA optimization factor scheme has achieved the best prediction effect under the fitting of the BP neural network model,followed by the SVR model,and finally the MLR model.The best fit schemes for CA and CA-PCA are the combination of factors when the input factors are 7 items.The combination of factors selected by the two factor optimization schemes of MI and MI-PCA is not satisfactory under the fitting of the three runoff forecasting models.(4)The model to obtain the optimal prediction result is the BP neural network model,and the input factor combination is CA-PCA7.The second is the SVR model with the input factor combination CA-PCA7.