基于随机森林算法的广州白云机场终端区雷暴潜势预报

THUNDERSTORM POTENTIAL PREDICTION BASED ON RANDOM FORESTS FOR THE TERMINAL AREA OF THE GUANGZHOU BAIYUN INTERNATIONAL AIRPORT

关键词:  灰色关联度  12小时随机森林分类模型  预报准确率AF  探测概率POD  MDRS  袋外错误率

Key words:  gray correlation degree  12 h forecast model based on random forests  forecast accuracy (AF)  detection probability (POD)  Massive Delay Response System (MDRS)  out of bag error rate

摘要:

利用2014—2015年夏季雷达资料获取雷暴发生有无数据，计算探空资料的对流指数与雷暴发生有无数据的灰色关联度，发现雷暴产生跟抬升凝结高度气压、850～700 hPa的温度和露点温度、风切变的关系最紧密。接着建立广州白云机场终端区内3类区域(离塔台中心8 km、50 km、100 km)的12小时随机森林分类模型，对不同区域的雷暴潜势进行预报和评估，发现终端区区域面积越大，雷暴发生样本比例越高，临界成功指数CSI、预报准确率AF、探测概率POD越来越高，虚假报警率FAR越来越低，表明预报出来的准确率越来越高。离塔台中心50 km和100 km区域的预报准确率AF和探测概率POD超过70%，对航空重要天气MDRS通报业务有指示作用。同时袋外错误率均低于1/3，随机森林算法的泛化性能好。

Abstract:

The present study used the thunderstorm occurrence data derived from the radar data from the Guangzhou Baiyun International Airport (ZGGG) terminal area and the convection index factors derived from the sounding data from nearby stations in the summers of 2014 and 2015 to calculate their gray correlations. The generation of thunderstorm is closely related to the pressure of the lifted condensation level (PLCL), temperature and dewpoint temperature from 850hPa to 700 hPa, as well as the wind shear. The 12 h forecast model based on random forests are established for three ranges (8 km, 50 km, and 100 km away from the center of the air traffic control tower) in the ZGGG terminal area. The evaluations of the thunderstorm potential prediction for different ranges show that the larger the terminal area, the higher the accuracy of the prediction, which can be demonstrated by higher critical success index (CSI), more forecast accuracy (AF) and detection probability (POD), as well as lower false alarm rate. The AF and POD with respect to ranges of 50 km and 100 km away from the center of the air traffic control tower are over 70%, which are useful indicators of major hazardous weather for aviation. The out of bag error rate being less than 1/3 confirms the fair generalization performance of random forests algorithm.