Research On Monthly Precipitation Prediction In North China Plain Based On Modified African Vultures Optimization Algorithm And Variational Modal Decomposition

Medium-term and long-term precipitation prediction is one of the important means to serve the macro-control of water resources and prevent future disasters in North China Plain.With the development of artificial intelligence technology,the machine learning model,as its branch,has achieved long-term development in improving the accuracy of medium and longterm precipitation prediction,showing its strong adaptability.Generally speaking,the prediction results obtained by directly inputting highly nonlinear influence characteristics into the model are often poor.Therefore,as a relatively mature time series noise reduction algorithm,Variational modal decomposition(VMD)is introduced to extract the unique frequency of precipitation or other meteorological feature series,which helps to improve the prediction effect.However,the inappropriate use of VMD technology will cause the illusion of "high precision prediction results",which cannot serve the actual prediction.Therefore,it is necessary to explore the differences between the two different decomposition methods of VMD algorithm-"full decomposition" and "stepwise decomposition" in constructing training samples and the possibility of serving the actual prediction.In addition,the selection of super parameters of machine learning model is also an important factor affecting the effect of medium and longterm precipitation prediction.In the study,a new intelligent optimization algorithm,the African Vultures Optimization Algorithm(AVOA),will be selected for further research.Based on orthogonal experimental design,a variety of combined AVOA algorithms will be formed from three factors: initial population,suboptimal vulture selection and parameter design.Two better combined optimization algorithms will be selected by using classical test function validation to obtain better model parameter optimization results.Finally,the coupling model formed by the above two strategies is applied to the case prediction of five representative sites in the North China Plain,and the main research results are as follows:(1)Two modified African vultures optimization algorithms(MAVOA1 and MAVOA2)are proposed.The vulture population is initialized by Zaslavskii chaotic map and Tent chaotic map,and the third vulture with poor fitness is regarded as the possible suboptimal vulture.Based on the orthogonal experimental method,18 combination algorithms are obtained by considering the four parameters of the original AVOA algorithm in a more precise step size.And two modified algorithms are selected based on the operation results on 23 test functions.(2)The conventional VMD full decomposition method is used to construct precipitation series training samples,and AVOA algorithm,MAVOA1 algorithm,and MAVOA2 algorithm are used to provide hyperparametric optimization means for LSSVM model,LSTM model,and RF model.The precipitation simulation results in the North China Plain show that the AVOALSSVM model has the best fitting effect,and the average training fitting evaluation is RMSE=5.6mm/month,MRE=71.0%,NSE=0.992;The average test fitting error is RMSE=6.1mm/month,MRE=73.0%,NSE=0.991.AVOA algorithm is more suitable for LSSVM model with less hyperparameters,while MAVOA2 is helpful to machine learning model with more hyperparameters.(3)Analyze the reason why the VMD full decomposition method cannot serve the actual precipitation prediction work through boundary effects.Based on the AVOA-LSSVM and MAVOA2-LSSVM models that performed well in the early stage,four stepwise decomposition techniques(SSD,FSD,SMSSD,and SMFSD)were used to construct training samples.Through comprehensive investigation of different VMD parameters α,conclusions were obtained: SMFSD-AVOA-LSSVM(α= 1000)is applicable to Huairou Station and Jingxian Station,with an average success rate of 59.8% and 54.5%;The monthly precipitation prediction model suitable for Zhengding station is SMFSD-AVOA-LSSVM(α=100),with an average forecast success rate of 61.2%;The monthly precipitation prediction model suitable for Hekou station is SMFSD-MAVOA2-LSSVM(α= 1000),with an average forecast success rate of 48.8%;The monthly precipitation prediction model suitable for Jiaozuo station is SMFSD-AVOALSSVM(α= 2000),with an average forecast success rate of 52.1%.In the horizontal comparison of stepwise decomposition sample technology,SMFSD is the most suitable tool for monthly precipitation prediction in the North China Plain.Among the prediction results of the five stations,Huairou Station(with a maximum error of-240.8mm,RMSE of 43.19 mm,NSE of 0.67,and MRE of 352.56%)and Jingxian Station(with a maximum error of-136.1mm,RMSE of 37.91 mm,NSE of 0.57,and MRE of 421.69%)perform a preferred overall effect,while Hekou Station has the worst effect(with a maximum error of-333.0mm,RMSE of64.10 mm,NSE of 0.32,and MRE of 191.58%).