Research On Application Of Multi-method Fusion Data-driven Model In Monthly Runoff Forecast

In recent years,the impact of human activities and climate change on runoff series has been further intensified,which makes runoff series show highly non-stationary and skewness characteristics,which brings challenges to runoff forecasting.Timely and accurate runoff forecast is of great scientific significance and practical value for solving the contradiction between water resources supply and demand and water resources planning and management.The Yellow River is the mother river of the Chinese people.General Secretary Xi Jinping regards high-quality development and ecological protection of the Yellow River basin as a national strategy.Medium and long term runoff forecast can provide certain scientific decision-making support for high-quality development of the Yellow River basin and solving the contradiction between supply and demand of water resources in the Yellow River Basin.Considering the variation characteristics of runoff,a monthly runoff data-driven model is established for the measured runoff from the main observation stations in the Yellow River Basin.By using four decomposition methods of EMD,EEMD,WD and VMD,two normal transformation methods of Box-Cox normal transformation and W-H inverse transformation,and four data-driven models of BP,SVM,RF and Elman,30 runoff prediction models were constructed,and the multi-step prediction system of the optimal model was established.The main conclusions are as follows:(1)The M-K trend test,Kendall rank correlation test,ordered cluster analysis and LeeHeghinian analysis were used to analyze the runoff evolution trend and abrupt change points of six stations in the Yellow River Basin.The results show that the runoff series of the six stations all have different degrees of non-stationarity.The variation points of Lanzhou station,Toudaoguai station,Zhangjiashan station,Zhuangtou station,Huaxian station and Lijin station occurred in 1985,1985,1996,1994,1968 and 1985,respectively.(2)Four single prediction models including BP neural network,SUPPORT vector machine(SVM),random forest(RF)and Elman neural network were constructed.The results show that the prediction accuracy of the four models varies greatly,and Elman and RF have better simulation results.The results of all models are not ideal in the validation period.(3)Sixteen hybrid forecasting models,including EMD-BP,EEMD-BP,WD-BP and VMD-BP,were constructed based on THE decomposition of EMD,EEMD,WD and VMD.The results show that the simulation results of the hybrid model are better than that of the single model,and the decomposition method can significantly improve the accuracy and stability of the model.The order of the comprehensive prediction effect in the validation period is VMD>WD>EEMD>EMD,indicating that VMD can decompose the original sequence into relatively stationary sub-sequences,make the sequence more fit the prediction model,and improve the prediction effect of the model.(4)Eight hybrid models of WH-BP,WH-SVM,WH-RF and WH-Elman were constructed based on W-H inverse transformation and Box-Cox transformation.The simulation results show that the hybrid model is better than the single model,and the normal transformation can significantly improve the accuracy and stability of the model.At the hydrological station of the Yellow River,the simulation effect of the optimal model based on W-H inversion is better than that of the optimal model based on Box-Cox transformation.In the hydrology station of the Yellow River tributary,the simulation effect of the optimal model based on Box-Cox transformation is better than that based on W-H inverse transformation.It shows that THE WH inverse transformation may be suitable for the model with high runoff,and the Box-Cox transformation is suitable for the model with low runoff.(5)Four hybrid models of VMD-WH-Elman,VMD-BC-Elman,WD-WH-Elman and WD-BC-Elman were constructed based on the optimized VMD-Elman,WD-Elman and W-H inverters transformation and box-Cox transformation,and the optimal model selected by each station was predicted by multi-step.The results show that VMD-WH-Elman and VMD-BCElman are superior to VMD-Elman.For the 30-month runoff prediction model of six hydrological stations,although the optimal model of each station is slightly different.However,in general,VMD-WH-Elman model constructed based on VMD,W-H transformation and Elman has the optimal accuracy and stability,R greater than 0.98,NSE greater than 0.95,NMSE less than 0.05,PBIAS less than 0.46,LME less than 0.87,and can still maintain high accuracy and stability in different prediction periods.(6)The App Designer platform based on MATLAB has developed a runoff prediction software,which visualizes runoff characteristics analysis and monthly runoff forecast,and can quickly and effectively analyze runoff characteristics and forecast runoff.