Runoff Prediction Based On Physics Informed LSTM Model

The runoff sequence has the characteristics of skewness,nonlinearity,and randomness.The runoff process includes deterministic components such as periodicity and trend,as well as random components such as underlying surface conditions,geographic location,and human activities.With a total length of 455.1km,the Jinghe River is the largest tributary of the Weihe River and a secondary tributary of the Yellow River.Medium-and long-term runoff forecasts can provide scientific decision-making basis for the development and utilization of water resources in the Jinghe River Basin and flood control for hydraulic projects.Considering the variation characteristics of the runoff sequence,this paper uses the daily runoff data of three controlled hydrological stations in the Jinghe River Basin,Qingyang Station,Yangjiaping Station and Zhangjiashan Station,and the meteorological data of four meteorological stations in Huanxian,Xifeng,Changwu and Pingliang.-Max Normalization(MM),Variational Mode Decomposition(VMD),Fuzzy C-Means Clustering(FCM),Lasso Regression,Pearson Correlation Coefficient Method and PMI Partial Mutual Information Method for 3 predictor screening methods,as well as BP neural network,Support vector regression machine,random forest and long-term memory neural network 4 models,established 20 kinds of daily runoff combined forecast models,carried out medium and longterm runoff forecast research,made comprehensive evaluation and optimization of the models according to 4 forecast error indicators,and obtained the following result:(1)Construct MM-Lasso and MM-Pearson prediction factor screening methods.For the same prediction model,the evaluation indicators of the three hydrological stations during the validation period show that the effect of MM-Lasso is better than that of MM-Pearson,indicating that Min-Max standardization is selected.Data processing and selection of Lasso regression to optimize predictors can effectively improve the forecasting effect of the model.(2)For the screening methods of VMD-MM-Lasso and VMD-MM-Pearson predictors,the comprehensive prediction effect is better than that of VMD-MM-Lasso;and the screening method using variational modal decomposition is better than the untreated one.The screening method has been adopted,which shows that the use of variational mode decomposition to decompose the data improves the prediction performance of the model.(3)The combination of fuzzy C-means clustering and PMI partial mutual information method constitutes a physical perception method.Compared with the runoff sequence processed by variational modal decomposition,the forecast effect in non-flood season runoff forecasting has been improved to a certain extent.(4)For the BP neural network model,the support vector regression machine model,the random forest model and the long-term memory neural network model,when the runoff sequence processing and predictor selection methods are the same,the four evaluation indicators in the verification period of the three hydrological stations show that The comprehensive prediction order of the four models is LSTM>RF>BP>SVM.It shows that the long-short-term memory neural network model can better utilize limited samples and has good generalization ability.Among the 20 combined forecast models of the three hydrological stations,the MMFCM-PMI-LSTM model showed relatively better forecast performance and stability in the daily runoff forecast of the Jinghe River Basin under the comprehensive comparison.During the verification period,the three hydrological The average relative error of the station is controlled at 15%,and the Nash efficiency coefficient is higher than 0.65,which has more obvious advantages.It is the preferred daily runoff forecast model for both Yangjiaping Station and Zhangjiashan Station,and it is the best model among the 20 models.,which can be used for daily runoff forecasting in the Jinghe River Basin.