利用双隐层BP网络进行雷暴潜势预报试验——以太原为例  

Experimental Research on Thunderstorm Forecasting with Double Hidden Layer BP Neural Network:Case Study on Taiyuan

作　　者：杨仲江[1] 蔡波[1,2] 刘旸[3] 

机构地区：[1]南京信息工程大学大气物理学院,南京210044 [2]沈阳军区空军气象中心,沈阳110015 [3]辽宁省人工影响天气办公室,沈阳110016

出　　处：《气象》2013年第3期377-382,共6页Meteorological Monthly

基　　金：公益性行业(气象)科研专项(GYHY200806014);江苏省高校优势学科建设工程资助项目(PAPD)共同资助

摘　　要：利用太原地区探空资料,结合闪电定位资料,采用神经网络法对太原地区雷暴天气进行潜势预报。选取与雷暴发生相关性较好的探空因子作为预报因子,对其进行归一化处理,输出采用两级分类,构建双隐层的BP网络,并应用独立样本进行预报检验。结果表明,在相同条件下,与单隐层BP网络相比,双隐层BP网络显示了其在解决分类问题上的优势;与多元统计回归法相比,双隐层BP网络获得更高的雷暴预报TS评分及更可靠的结果,显示出神经网络良好的非线性问题处理能力。并且对雷暴预报结果的规律进行了分析与讨论,说明探空因子与雷暴的发生有着密切的联系。A neural network scheme to do a multivariate analysis for forecasting the occurrence of thunderstorm in TaiYuan is presented by using sounding data and lightning location system data. Well correlated sounding factors are selected as the predictors, then all the input factors are normalized, and output data are adopted to two-stage category so that the BP network with double hidden layers has been established and the independent samples can be tested in it. The results indicate that, in the same condition, compared with single hidden layer BP network, the double hidden layer BP network shows its advantage on solving classification problem. Compared with multivariate statistics regression algorithm, the neural network algorithm obtains higher thunderstorm forecasting TS score and more reliable results, showing good nonlinear processing ability in the thunderstorm forecasts based on sounding data. And then the rules of thunderstorm forecast results are analyzed and discussed, showing that sounding factors have a close connection with the occurrence of thunderstorm.

关 键 词：神经网络 BP网络 雷暴预报 探空资料 

分 类 号：P456[天文地球—大气科学及气象学]