| Scientific and reasonable runoff prediction can provide a good decision-making basis for the prevention and control of flood and drought disasters in the basin and the scientific allocation of water resources.However,in the context of global climate change,the steady-state assumption of hydrological system of the basin is no longer valid,which makes accurate runoff prediction more and more difficult.Therefore,based on machine learning theory and hydrological teleconnection analysis,this thesis proposed a combined monthly runoff prediction model that integrates large-scale climate factors,and selected three representative hydrological stations in Heihe,Weihe,and Pearl River basins in different climate zones in China.The monthly runoff data is used as the research object to test the actual prediction effect of the combined prediction model and various conventional prediction models.The main research content and results of the thesis are as follows:（1）The time-frequency decomposition method is used to reveal the multi-scale time information contained in the non-stationary original runoff sequence,that is,the runoff is serialized into a series of relatively stable intrinsic mode function（IMF）series through the improved adaptive noise complete ensemble empirical mode decomposition（ICEEMADAN）,and then use variational mode decomposition（VMD）to perform secondary decomposition on the high-frequency IMF1 series.Finally,the predictors of ICEEMDAN-VMD hybrid decomposition subsequences were selected by partial autocorrelation coefficient（PACF）and SMAC-Ada Boost,and used as input variables of the machine learning model.Comparing the prediction results of each model shows that the predictors selected by SMAC-Ada Boost are more conducive to the machine learning model to capture the general law of runoff change.（2）Teleconnection regression model and runoff time series analysis model are two types of widely used data-driven models in hydrological forecasting.Based on the mutual information theory（MI）and the SMAC-Ada Boost method,this paper explores the influence of large-scale climate factors on the teleconnection of watershed runoff,in addition,the predictors which have significant effects on each hydrological station were selected from 130 climate factors,and the regression model between it and the predicted runoff was constructed.Combining with the time series analysis method,the chaos-echo state network model（C-ESN）,the seasonal differential autoregressive moving average model（SARIMA）and the CNN-LSTM neural network model（CLNN）were established,respectively.Compared with the teleconnection regression model,the CLNN runoff time series analysis model has the best effect.（3）Conventional runoff time-series analysis models usually cannot accurately capture some anomalies in watersheds affected by abnormal atmospheric circulation.In order to strengthen the physical significance of the runoff prediction model,The combined monthly runoff prediction model（T-IVCL）incorporating large-scale climate factors was proposed.The T-IVCL model predicts runoff by combining the CNN-LSTM network layer that extracts the feature information of the ICEEMDAN-VMD and the BP network layer that extracts the climate factor information.In this way,the dynamic information of runoff time series and the teleconnection information of climate factors can be considered at the same time.The Nash efficiency coefficients（NSE）of T-IVCL model for monthly runoff prediction of Yingluoxia,Xianyang and Boluo hydrological stations are 0.961,0.902 and 0.913,respectively.Compared with the IVCL model that only considers the time-frequency decomposition information of runoff,the NSE of the combined model increased by 0.2%,3.1%,and 2.5%,respectively.（4）A runoff interval prediction model based on Gaussian process regression（GPR）was constructed.GPR can not only obtain reliable runoff prediction intervals of each model,which can provide further decision-making basis for scientific management of basin water resources,but also quantify the stability（SM）of each model in the runoff prediction process of this study.Compared with the IVCL model,the SM of T-IVCL model for the monthly runoff prediction results at Yingluoxia,Xianyang and Boluo hydrological stations decreased by 2.3%,7.0%and 17.5%,respectively.The results of point estimation and interval estimation of T-IVCL model show that the data-driven model integrating large-scale climate factor information is more adaptable to the monthly runoff prediction under environmental change. |
