Abstract:Based on the data of the NCEP global reanalysis,with a resolution of 1°×1° and cloud-to-ground lightning in Shaoyang from June to August 2008 to 2010,an SVM classification model of thunderstorm forecast was set up.First,the ground flashes data for 6 hours corresponding to the NCEP data were counted.The flashes were greater than or equal to 3,which is defined as a thunderstorm.A non-thunderstorm is defined as when the flash is zero.If the number was greater than 0 or less than 3,it was not included in the statistical samples.In addition,if the flashes were detected by the lightning location system,but the thunderstorm days were not recorded at the meteorological observation station,then the data were not included on the model samples.According to these criteria,1 007 samples were counted,286 of which were thunderstorm samples,and 721 were non-thunderstorm samples.80% of the samples were used as the training samples,and the remaining 20% were used as the test samples.Next,27 parameters were extracted from the NCEP global reanalysis data on 1°×1° degree grids,and 21 convection parameters were calculated,of which a total of 48 physical quantities were used to set up a thunderstorm-forecasting model according to the method of SVM classification.The correlation between 48 physical quantities and the occurrence of thunderstorm of training samples was calculated,then 11 significant predictors greater than 0.3 were selected.Before the forecasting model was established,all parameters were normalized.By repeatedly changing the parameter values,the highest classification accuracy of the parameters c and g were finally selected as the model parameters.When the two parameters were determined,the potential thunderstorm forecast model was set up.The test results show that the prediction accuracy of the SVM model was 86.21%,false alarm rate was 15.25%,and missing rate was 13.79%.The scope and step length were manually set during the parameter optimization,yet after several repetitions some modeling errors were still present,which led to the test results differing.In order to validate the stability of the model,70% (90%) of the modeling samples were re-selected and verified with optimal parameters.It was found that there was no significant difference in accuracy among the three kinds of model samples,which illustrates the stability and reliability of the forecast model,and the model can forecast the thunderstorm in 6 hours.Logistic regression discrimination and Bayesian discrimination are the two other forecast methods used for thunderstorms.Through comparison of the three forecast methods,it was found that the forecast model based on SVM gave the best prediction performance with TSS,and the true skill statistic was 0.79.The rates of the false alarm and pseudo fault report were lower than when using the other two methods.Therefore,it can be seen that the forecast model built using the SVM method is able to provide a certain reference value for potential thunderstorm prediction in the Shaoyang area,with a higher accuracy rate of distinguishing thunderstorms and non-thunderstorm.In other words,the prediction model can provide a reference value for potential thunderstorm forecast in the Shaoyang area.