Study On Runoff Prediction Based On Modified Empirical Mode Decomposition And Ensemble Learning

In the wake of development in computer and artificial intelligence,its related technologies have penetrated into the various walks of life.In the world of hydrology and water resources,artificial intelligence can analyze and handle massive hydrology and water resources data.As an important component of hydrology,the calculation and prediction of runoff are important task of water conservancy construction,and also an important orientation of computer application.The generation of runoff is influenced by various factors,which is a process with high complexity.The aim of this thesis is to use the AI-related technology to study the runoff evolution law,explore the factors affecting the runoff generation process,and on this basis to make accurate and stable prediction.Based on the above purpose,the monthly runoff data from 1950-2020 of four stations in the lower reaches of the Yellow River were selected for the study.And the runoff influencing factors of four hydrological stations are analyzed.At the same time,the monthly runoff prediction model was established by combining with the neural network and ensemble learning,which provides reliable technical support for the region’s mid and long-term runoff forecast.The main contents of this thesis are as follows:(1)The distribution,tendency,mutagenicity and other features of the monthly runoff data from four stations were analyzed in a holistic manner by using hydrological statistical tools and mathematical methods.Meanwhile,the cause of its generation was analyzed in the light of historical climate.(2)Complementary ensemble empirical mode decomposition(CEEMD)was used to decompose the runoff,and its shortcomings were analyzed.In view of the limitations of CEEMD in processing runoff data,three improved algorithms are proposed as follows:CEEMD with high frequency components removed(RCEEMD),CEEMD with adjacent frequency terms were refactoring(ICEEMD),CEEMD was reconstructed after removing the high frequency components(MCEEMD).The decomposition algorithm was combined with the minimum gate unit recurrent neural network(MGU)to establish the runoff prediction model.The prediction results were verified,and the MCEEMD-MGU model reached Grade B of the standard of hydrological prediction models,and the accuracy of it’s prediction result was also improved in different degrees,indicating that MCEEMD could process runoff data more effectively.(3)The theory of comentropy and principal component analysis(MI-PCA)was combined to explore the meteorological factors affecting runoff.The results show that sunspot,northern hemisphere,Eurasia and Pacific Ocean are the main factors affecting the generation of runoff.Based on the MI-PCA method,a hybrid multi-feature runoff prediction model combining MCEEMD and modified Stacking(M-Stacking)ensemble learning was proposed.The weight coefficients of different individual learning algorithms in the M-Stacking ensemble learning were given by the result error,and GAP-5 fold cross validation was used to partition the data.Results show that the M-Stacking ensemble learning model reaches the Grade A of the hydrological prediction model standard,and the qualified rate and accuracy were improved compared with other advanced models.Furthermore,the runoff prediction error in flood season of the ensemble learning is also reduced compared with other models.In conclusion,the MStacking ensemble learning model has a good performance in runoff prediction with high stability and early warning ability,which makes up for the defects of the traditional single prediction model,and also provides an effective means for the medium and long term prediction of regional runoff.