Simulation And Mid-long Term Forecast Of Monthly Runoff In The Source Region Of The Yellow River Based On Machine Learning

Mid-long term streamflow forecast is of great significance for water resources management and it is also an important basis for reservoir operation.The source area of the Yellow River(SYRY)is the major area for runoff generation in the Yellow River Basin.The research on the mid-long term runoff forecastis is important for water resources management in the SYRY.Compared with the traditional hydrological model,the machine learning model can simulate and forecast the runoff more efficiently.Therefore,the research on the application of machine learning method in runoff simulation and prediction is also of great practical value.Based on two forecasting factor selection methods and two precipitation forecasting methods,four mid-long term precipitation forecasting models were constructed.The comparison of the forecast results of the four models shows that the prediction accuracy of the Mutual InformationNARX Neural Network model is higher than that of the other models,and the relative error of the forecast of the precipitation in flood season from May to September is less than 20%.And based on the Mutual Information method and the NARX Neural Network,the mid-long term forecast model of monthly potential evapotranspiration and temperature in flood season was also developed.The relative error of the model for the forecast of potential evapotranspiration was less than 5%,and the relative error of temperature forecast is less than 15%.The monthly runoff variation the study basin upstream the Tangnaihai Hydrological Station in the SYRY from 1965 to 2015 was analyzed.Using three machine learning models,namely Support Vector Machine(SVM),Reverse Artificial Neural Network(BP-ANN)and Random Forest(RF),the monthly runoff were simulated.The accuracies of different models were compared and analyzed,and the machine learning model were also compared with the conceptual hydrological model ABCD-CR and the tank model considering the characteristics of cold regions.The results show that the performance of the SVM and BPANN model is better than that of RF and the two conceptual hydrological models.The BPANN model is better than the SVM model in the simulation of runoff in the flood season.Therefore,the BPANNmodel can be used as the optimal model for runoff simulation in the flood season of SYRY.Combined with NARX-ANN and BPANN,a mid-long term runoff forecast model for SYRY in the flood season was developed,and was validated based on the observations.The results show that the mid-long term runoff forecast model can accurately predict the runoff in the flood season in the source region of the Yellow River.The NSE of the result of the runoff in flood season is 0.64 and the relative error was 20% in the training period(1965-2000),the NSE is0.59 and the relative error was 18% in the testing period.And the mid-long forecast model can accurately predict the flow peaks in the flooding year.The results also show that the selection of precipitation predictors has great influence on the forecast results.When the number of predictors increased,the accuracy of the runoff forecast decreased.The results of this study can provide an effective tool for reservoir management in the upper reaches of the Yellow River,and were potentially useful for water resources management in the SYRY.