基于XGBoost修正的融合神经网络在降水量预测中的应用OA

When dealing with nonlinear and non-stationary precipitation series,traditional precipitation prediction models are often difficult to fully capture complex time series characteristics,resulting in systematic deviations in prediction results.Therefore,this study explored the ability of extreme gradient boosting(XGBoost)to identify and correct precipitation prediction residuals.The monthly meteorological data of five meteorological stations in the Sanjiangyuan area were taken as the research case.Firstly,Kendall's rank correlation analysis was used to objectively screen out the first six meteorological factors with a high correlation with precipitation.Then,the ensemble empirical mode decomposition(EEMD)was used to decompose the meteorological elements at multiple scales,and the modal function(IMF)components of each meteorological element were extracted.Secondly,these components were used as the input of long short-term memory(LSTM)and a convolutional neural network(CNN),respectively.The complex features of the precipitation sequence were learned by the deep learning model,and the prediction sequence with a large residual pattern was output.On this basis,XGBoost was innovatively introduced to identify and correct the residual model of monthly precipitation prediction of LSTM and CNN,and the optimization and reconstruction of the original prediction results were realized.The LSTM-XGBoost and CNN-XGBoost fusion models were constructed for precipitation prediction.The results show that after XGBoost residual correction,the precipitation prediction performance of Wudaoliang Station is improved,and the RMSE,MAE,and R2 are improved to varying degrees.The increase of R2 is 2.96%(LSTM)and 3.88%(CNN),respectively,and the distribution of precipitation prediction error is more concentrated.The systematic prediction deviation of the basic model is effectively corrected,and the stability and reliability of the prediction results are significantly enhanced.In addition,cross-site verification shows that the Kling-Gupta efficiency coefficient(KGE)of different meteorological stations has different degrees of improvement,and the residual correction mechanism of the constructed fusion model is universal.Based on the above,it is revealed that XGBoost effectively optimizes the prediction residuals by capturing the nonlinear residual mode of a single deep learning model,which eliminates the prediction error to a certain extent,improves the prediction accuracy of precipitation and the prediction stability of the model as a whole,establishes an interpretable residual correction mechanism,and provides a new perspective for understanding the source and characteristics of the prediction bias of the model.This study provides a new technical method for future precipitation prediction and precipitation risk assessment and has important theoretical value and practical significance.