一种基于BDC雷暴预报的最优子集方法研究

Research on an optimal subset method based on BDC thunderstorm forecast

二进制粒子群-贝叶斯判别准则(Binary Particle Swarm Optimization-Bayes Discriminatory Criterion,BPSO-BDC)方法,能够使用二进制粒子群(Binary Particle Swarm Optimization,BPSO)算法智能地选取出贝叶斯判别准则(Bayes Discriminatory Criterion,BDC)雷暴预报模型的最优子集,克服了BDC在因子选择时的缺点。为了建立BDC雷暴最优模型,利用2010—2014年T511数值预报产品和单站观测资料,对漳州、义乌、乐东三站BDC雷暴预报模型进行研究。通过选取适应度函数,提出了BPSO搜索BDC模型最优子集的计算方法,得到了三站的最优子集模型,并与BDC和逐步判别模型进行对比。结果表明:在24 h的雷暴预报结果中,BPSO-BDC模型的平均TS评分达到了0.697,空报率为0.256,漏报率为0.048,在48 h的雷暴预报结果中,BPSO-BDC模型的平均TS评分达到了0.418,空报率为0.222。BPSO-BDC模型的预报结果明显优于BDC和逐步判别模型。对BPSO-BDC模型进行稳定性检验,TS评分都在0.21~0.35之间,变化幅度较小且处于较高水平。说明BPSO-BDC方法预报效果明显优于BDC和逐步判别方法,且具有良好的稳定性。

A new thunderstorm prediction method, namely, Binary Particle Swarm Optimization-Bayes Discriminatory Criterion（BPSO-BDC） was proposed, which can use Binary Particle Swarm Optimization（BPSO） algorithm to automatically screen the optimal subset of Bayes Discriminatory Criterion（BDC） model, overcoming the shortcomings of BDC model in factor selection. In order to establish the optimal model of BDC thunderstorm, the BDC thunderstorm forecasting model for three stations at Zhangzhou, Yiwu and Ledong was studied, by using T511 numeral prediction product and single-station observation data of 2010-2014. By selecting the fitness function, the optimal subset of BPSO-BDC model is proposed, and the optimal subset model for the three stations is obtained and compared with BDC and stepwise discriminant model. Results show that for thunderstorm prediction through equation established by BPSO-BDC method, in the 24 h thunderstorm forecast, its mean value of threat score reaches 0.697, mean value of the false alarm rate is 0.256, and mean value of the missing alarm rate is0.048. In the 48 h thunderstorm forecast, its mean value of threat score reaches 0.418, mean value of the false alarm rate is 0.222. Results of BPSO-BDC model is obviously better than BDC and stepwise discriminant model. The results of the BPSO-BDC model showed that the TS scores were between 0.21 and 0.35, and the variation range was small and at a high level. Prediction effect of BPSO-BDC model is obviously better than BDC and stepwise discriminant, with good stability and prediction capability.