Research Of Multi-source Meteorological Data Based On Machine Learning

Apart from destroying agricultural production,short-term heavy rainfall may also seriously threaten the normal operation of social,political and economic life.Therefore,shortterm precipitation forecasting,a study beneficial to the country and the people,is worthy of further research.This paper is devoted to applying machine learning methods to precipitation prediction,and two precipitation prediction studies are carried out for surface meteorological observation data and radar meteorological observation data respectively.The first is the research based on surface observation data: To address the problem that knearest neighbor algorithm ignore the influence of different attributes on the classification results when dealing with multi-dimensional data classification,a new k-nearest neighbor algorithm based on attribute significance is proposed to improve the classification performance of classifiers.Using pressure,wind direction,wind speed,temperature and relative humidity as the feature of the sample and the observed precipitation data as the class,a weight is set to each attribute according to the cohesion and heterogeneity of data distribution of same category.Thus,the neighborhood search is implemented by weighted Euclidean distance to achieve optimal classification.The results show that the new precipitation model performs better: the accuracy,TS score and positive sample summary rate of forecast result are increased;the standard error and false negative rate of precipitation prediction are reduced.The second is the research based on radar observation data: aiming at the situation that the existing models for forecasting precipitation are unsatisfactory in forecast errors and the generalization capability,this paper combines the advantages of CNN and RNN.Two precipitation prediction models,Conv3D-Bi LSTM and Conv3D-Bi GRU,are designed and constructed in this paper.Experimental results show that the two models have their own advantages in the field of radar short-term precipitation prediction.In order to improve the prediction effect of the model,this article introduces the stacking integration strategy to merge Conv3D-Bi LSTM and Conv3D-Bi GRU,and then a precipitation prediction model based on integrated deep neural network is proposed.Experimental results show that the new model performs better in prediction accuracy and convergence effect.