The Research Of Optimal Subset Technology Based On Bayesian Forecast Model

Thunderstorms containing huge unstable energy to produce wind,rain,lightning and other severe weather,posing a great threat to human activities and people's life safety and property safety,has always been the focus of the weather forecast service.In this paper,thunderstorm generated by strong convection was taken as the object of study,and a new method of thunderstorm forecasting was proposed,the binary particle swarm optimizatio nNa?ve Bayesian classifier(BPSO-NBC)and the binary particle swarm optimization-Ba yes discriminatory criterion(BPSO-BDC)method.Using BPSO algorithm,combined with the NBC and the BDC,we can automatically select the optimal subset of the NBC and BDC models.It not only makes the tedious NBC and BDC model's forecasting factors easier and more convenient,but also makes their prediction results optimal for all forecast subsets.At the same time,the performance differences between BPSO-NBC,BPSO-BDC thunderstor m model and the classical models such as stepwise discriminant analysis and K nearest neighbor nonparametric regression model were analyzed and compared.And the stability of various models was analyzed.The main contents are as follows:First,NBC and BDC thunderstorm forecasting models were established.Using 2010~2014 year by year August T511L60 medium range numerical forecast product,27 strong convection indexes,such as K index and S index,were calculated.With Suixi station,Lingshui station,Ledong station observation data,an interpretation model was formed to establish the required training set and test set.In view of the frequency of thunderstor m occurrence,after using the minimum historical value of the strong convection index to eliminate the false prediction,the frequency of thunderstorms was increased by 2-3 times in training concentration.The NBC and BDC models were selected by Fisher discrimi na nt criteria.Among them,the NBC model directly used the Fisher criterion to rank the strong convection index,and the larger one was used as the prediction factor;The BDC model was firstly screened by t test,when the t test was not satisfied with the significance test criterion,the Fisher criterion was adopted to select the forecasting factor.Based on the data set collected above,the NBC and BDC thunderstorm forecasting models of three stations were established by NBC and BDC methods respectively.2010~2013 years of data was used for the fitting analysis,and the 2014 data was tested.The fitting results show that the average TS score of three stations in NBC model and BDC model is above 0.20,and within 24 h,the fitting effect is the best.The test results show that the average TS score of three stations in NBC model is above 0.18,which is lower than the fitting level.With the increase of the forecast time,the overall forecast level has declined.Then,BPSO-NBC and BPSO-BDC thunderstorm forecasting models were established.By using the BPSO algorithm,the optimum subset of the NBC and BDC model was selected by designing the fitness factor,and the thunderstorm forecasting model of BPSO-NBC and BPSO-BDC was established.The fitting results show that the average TS score of the three stations in the BPSO-NBC model is above 0.35,and the average TS score of the three stations in the BPSO-BDC model is above 0.29,which is obviously higher than that of the NBC and BDC models.The test results show that the average TS score of NBC model is above 0.27,and the average TS score of BDC model is above 0.19,which is lower than the fitting level.But the average results of three stations in BPSO-NBC and BPSO-BDC models are better than those of NBC and BDC model,and the effect is obvious after 24 h.Finally,several Bayesian models were compared with stepwise discriminant and K nearest neighbor nonparametric models.From the average TS score,BPSO-NBC,stepwise discriminant,NBC,BPSO-BDC,K nearest neighbor,BDC,the TS scores of 6 models are 0.276,0.210,0.202,0.192,0.187,0.155,respectively,and BPSO-NBC model is the best.Their rate of empty reporting is similar,and the rate of missed reporting is the same as that of TS.From the results of stability test,the stability of the BPSO-NBC model is best,three TS scores were stable at 0.2,BDC,K and NBC model station TS score is above 0.18,the stability of BPSO-BDC model was more than 0.17,stepwise discriminant model of TS score was more than 0.12.The result shows that the NBC model and BDC model can significantly improve the thunderstorm prediction level of the Bayesian model through the BPSO algorithm.It also provides new ideas and methods for single station thunderstorm forecasting and other optimization algorithms and models that can not automatically select factors.