Medium And Long-term Runoff Forecast Of The Yellow River Basin Based On Ensemble Learning Algorithm

Medium and long-term runoff prediction is one of the key research contents of hydrologists.Accurate runoff prediction can provide scientific decision-making for water resources planning and management,reservoir optimal operation and social sustainable development.For a long time,the hydrological models have been the powerful tool for carrying out runoff prediction research.However,the hydrological cycle is a dynamic and complex system.Under the dual impact of climate change and human activities,the hydrological process of the basin has undergone profound changes,the uncertainty has increased significantly,and the non-stationarity and non-linearity of the runoff series are increasingly intensified.The limitations of the existing hydrological models in practical application are gradually highlighted,and the runoff prediction is facing new challenges.How to fully grasp the current hydrological process and improve the accuracy of runoff prediction has become a research hotspot.The Yellow River is the second largest river in China and an important freshwater resource in the north of China.With the continuous development of economy and the rapid advancement of urbanization,climate change and high-intensity human activities have significantly disturbed the hydrological situation in the basin.The underlying surface of the basin has changed,and the frequency and severity of extreme hydrological events have increased.The complexity and non-stationarity of runoff series have become more prominent.It is of great practical significance to carry out the research on medium and long-term runoff forecasting to realize ecological protection and high-quality development in the Yellow River Basin.Therefore,the monthly runoff series of six hydrological stations in the main stream of the Yellow River is taken as the research object.The elastic net,Bagging and Boosting,and variational mode decomposition method are used to construct the monthly runoff forecasting model based on the decomposition-ensemble framework,exploring medium and long-term runoff forecasting methods to improve runoff forecasting accuracy and provide scientific basis for the development and utilization of water resources,flood control and drought relief in the basin.The main contents and achievements of this study include:(1)The M-K test and Hurst index are used to analyze the trend and possible future trend of runoff of each hydrological station,and the M-K test and Pettitt test are used to identify the year of abrupt change point of runoff of each hydrological station,so as to reveal its evolution trend and law.The results show that the runoff of each hydrological station shows a decreasing trend,and the trend is sustainable.Except for Tangnaihai station,the runoff of the other five hydrological stations decreased significantly,and the trend is sustainable.The runoff mutation characteristic of each hydrological station is significant,and the abrupt change points are mostly concentrated in the 1980s～1990s.(2)The Elastic Net(EN)is used to screen the forecast factors,which determines the input of the model.And the correlation coefficient method is compared to explore the potential of EN in screening key forecast factors.Aiming at the problem that the prediction ability of individual learning is limited,the random forest model(RF)is built based on the Bagging ensemble learning strategy,and the gradient boosting decision tree model(GBDT),the extreme gradient boosting model(XGBoost),and the light gradient boosting machine model(Light GBM)are built based on the Boosting ensemble learning strategy to comprehensively capture the change characteristic of the runoff series.The results show that the EN can reduce the interference of redundant or irrelevant factors on the prediction model while ensuring the group effect,and achieve the purpose of dimensionality reduction,which has a positive effect on the prediction and can make the prediction model get better results.In general,the prediction results of the four ensemble learning models in the training period generally meet the prediction accuracy requirements,but the prediction results in the testing period have different degrees of error,and only the prediction results of Lanzhou station are reliable.(3)The ensemble learning model has the ability to pay attention to samples with high error,but it will also lead to sensitivity to the noise of runoff series under changing environment,and its generalization ability will be reduced.Therefore,the variational mode decomposition(VMD)technology is used to obtain sub-series with different frequencies to mitigate the impact of noise on the prediction,and the ensemble learning models integrated the decomposition technology is constructed to realize the monthly runoff prediction under different periods.The results show that,compared with the four single ensemble learning models,the decomposition-ensemble prediction models can more accurately grasp the change rules of runoff series and capture effective information,and use the information to accurately predict.With the increase of the prediction period,the forecast accuracy of the decompositionensemble prediction models has decreased,but it can still provide reliable runoff prediction.Based on evaluation indicators,the prediction effectiveness of the decomposition-ensemble prediction models is ranked as follows: VMD-EN-Light GBM>VMD-EN-GBDT>VMD-ENXGBoost>VMD-EN-RF.(4)On the basis of deterministic prediction,the non-parametric estimation method,kernel density estimation(KDE),is used to describe the distribution of prediction error,so as to describe the comprehensive uncertainty of the decomposition-ensemble prediction models and realize the probability interval prediction.The results show that the coverage of the four decomposition-ensemble prediction models at 90% confidence level is not less than 90%.The prediction intervals contain reliable risk information,which can accurately quantify the uncertainty of runoff prediction.The prediction interval width of the decomposition-ensemble prediction models for the downstream is larger than that of the upstream and midstream,indicating that the uncertainty is larger,which is closely related to human activities.Based on evaluation indicators,the VMD-EN-Light GBM model has the best prediction effect,achieving ideal coverage with low uncertainty and providing reliable interval prediction.(5)The applicability of CMIP6 in the Yellow River Basin is evaluated by using the optimal decomposition-ensemble prediction model(VMD-EN-Light GBM)analyzed in the previous article.In order to reduce the impact of the limitations,the data of climate models of CMIP6 are corrected.The results show that the accuracy of the data after the bias correction is significantly improved compared to that before bias correction,which can accurately describe the monthly precipitation and monthly average temperature of the Yellow River Basin.The accuracy of runoff prediction is reduced when the revised data are used as inputs,but it can still provide reliable runoff prediction,which is not significantly different from the prediction results based on the measured data of meteorological stations,indicating that the runoff prediction results obtained by the data of climate models driven decompositionensemble prediction model are reliable.