基于支持向量机的雷暴大风识别方法

基于北京市观象台雷达基数据和加密自动气象站数据，利用支持向量机算法建立了雷暴大风天气的有效识别模型。首先确立了9个用于识别雷暴大风的预报因子:回波顶高、最大反射率因子、最大反射率因子所在高度、垂直积分液态水含量、垂直积分液态水含量随时间变率、垂直积分液态水含量密度、雷暴大风发生前最大反射率因子下降高度、风暴移动速度、速度谱宽。通过计算各预报因子在大风和非大风样本中的概率分布，得到对应的各项预报因子雷暴大风的隶属度，利用得到的隶属度函数对样本进行归一化处理。确立核函数和模型参数，利用支持向量机建立雷暴大风天气的提前识别和临近预警模型。通过对北京2017年7月7日飑线和2012年5月19日块状回波引起的灾害大风典型个例的识别效果检验，得到两个个例预测的命中率、误判率和临界成功指数分别为92.0%，22.1%，73.0%和99.1%，40.5%，59.2%，对于提高雷暴大风预警预报的准确率有一定帮助。

Thunderstorm Gale Identification Method Based on Support Vector Machine

A thunderstorm gale recognition model is established using support vector machine based on data of radar and automatic weather stations from Beijing Weather Observatory. Firstly, 18 thunderstorms in Beijing during 2010-2014 are analyzed quantitatively in terms of the statistical method and 9 forecast factors are selected, i.e., the height of the echo top, the maximum albedo, the height of the maximum reflectivity, the total vertical liquid water content, the time rate change of total vertical liquid water content, the total vertical liquid water content density, the height of the maximum reflectivity factor, the storm moving speed and the width of the velocity spectrum. 451 non-high wind samples and 425 high wind samples are selected by matching the time and place of automatic weather stations with the value of the quantitative index of the PUP storm monomer recognition product in all the cases. Secondly, the probability distribution of prediction factors in the wind and non-wind samples are calculated, and relationships with corresponding forecast factors are obtained, and then sample data are normalized by the obtained membership function. Finally, the kernel function and model parameters are established, and the thunderstorm gale recognition model is established using support vector machine. Two typical cases in Beijing are analyzed and tested, one caused by a line thunderstorm which happened on 7 July 2017, and the other caused by an isolated single-cell storm which happened on 19 May 2012. Results show that the identified wind range is consistent with reality, and the hit rate, the false alarm rate and the critical success index are 92.0%, 22.1%, 73.0% and 99.1%, 40.5%, 59.2%, respectively. It will help to improve the accuracy of thunderstorm gale warning and forecasting. However, sometimes multiple thunderstorm cells can be misjudged as one according to these forecast factors. In this case, it is necessary for forecasters to conduct manual intervention in combination with the overall radar base reflectivity and weather conditions, to reduce the misjudgment rate of gale. In the future, long time series radar data should be used to carry out "large sample census" research and an automatic thunderstorm identification system based on weather radar can be established.