遗传—神经网络集合预报方法在广西热带气旋降水预报中的应用  

An genetic neural network ensemble prediction model for tropical cyclones rainfall in Guangxi

作　　者：黄小燕[1] 赵华生[1] 黄颖[1] 林开平[2] 何立[1] 

机构地区：[1]广西区气象减灾研究所,广西南宁530022 [2]广西区气象台,广西南宁530022

出　　处：《自然灾害学报》2017年第6期184-196,共13页Journal of Natural Disasters

基　　金：国家自然科学基金(41575051;61562008);广西区气象局气象科研计划重点项目(桂气科2016Z03);广西青年基金(2014GXNSFBA118211)~~

摘　　要：利用多维尺度分析可以从一系列数据集的相似性信息中发掘其中的潜在结构信息,并通过样本点间的相似度构建相似矩阵,再将相似矩阵映射到低维欧氏距离空间获取新的特征的能力。论文以1980—2015年共36年的广西热带气旋逐日降水量为基础,综合考虑热带气旋降水的数值预报产品物理量预报因子,采用多维尺度分析的预报因子信息数据挖掘技术,以进化计算的遗传算法,生成期望输出相同的多个神经网络个体,建立了一种新的非线性人工智能集合预报模型,进行遗传—神经网络的广西热带气旋降水集合预报模型研究。遗传—神经网络的集成个体的输入因子是通过选出相关程度较高的数值预报产品的物理量场格点因子,同时网络的输出是通过多维尺度降维方法对初选预报因子群进行合理的降维处理来实现的。通过对2011—2015年影响广西的22个热带气旋共94个独立样本的试验预报的统计结果表明,对于暴雨以上量级的预报,广西89站的预报平均TS评分达到了0.3;同区域同样本的对比分析表明,新预报模型5年的TS评分比欧洲细网格的TS评分提高了15%以上;进一步对大量级的降雨落区的预报对比分析表明,新方案的预报效果比欧洲中心的预报更理想。We developed a novel nonlinear artificial intelligence ensemble prediction （NAIEP） model based on multiple neural networks with identical expected output created by using the genetic algorithm （GA） of evolutionary computation. We extracted the potential structure information from the similarity information of series of data sets by multidimensional scaling （MDS） analysis, and the similarity matrix is constructed by the similarity between the sample points. Then the similarity matrix is mapped to the low dimensional Euclidean distance space to obtain the new feature. We set up the Genetic Neural Network Ensemble Prediction （GNNEP） model based on data of the daily precipitation of tropical cyclones （TC） that occurred in Guangxi in the period 1980-2015 for 36 years. The predictors were selected by MDS in the Numerical Weather Prediction products to predict the TC precipitation for each station. The input factors of GA neural network are selected by selecting the physical quantities of the products with higher degree of correlation. At the same time, the output of the network is realized by the dimensionality reduction of the primary forecasting factor group by MDS. The results of the experimental forecast of 94 independent samples of 22 tropical cyclones in Guangxi from 2011 to 2015 show that the average TS score of the 89 stations in Guangxi is 0.3 for the forecast of the magnitude above the heavy rain. The comparative analysis shows that the new forecast model for 5 years TS score than the European fine grid TS score increased by 15% or more under the condition of the same region and samples, and further large magnitude of rainfall area forecast comparison analysis shows that the forecasting effect of the new scheme is more ideal than by the European Center.

关 键 词：热带气旋 降水预报 数值预报产品 多维尺度 遗传-神经网络集合预报 

分 类 号：P457[天文地球—大气科学及气象学] X43[环境科学与工程—灾害防治]