Abstract:

Based on the air quality data of five indices in 2010 for 78 main cities of China, the research calculated the comprehensive score of urban air quality, selected ten out of 48 variables describing the climate, topography, urban development and environment management of these cities with multivariate linear regression analysis, and quantified their contribution to the urban air quality. Based on the comprehensive score of urban air quality, the authors used a stratified random sample of 30 from the 78 cities, as a training sample, to construct a radial basis function network (RBFN) model, which was used to simulate air quality of 173 main cities in China based on the natural and social-economic features, and environmental management of the cities. The results indicated that the average saturation vapor pressure, built-up urban area, elevation range, and the percentage of industry in GDP as four major dominants of urban air quality, accounting for the variation by 14.7%, 12.8%, 8.8% and 7.2%, respectively. This study broke the limitation of most previous air quality assessment models, which merely took air pollutants and meteorological factors as input. The result showed a high accuracy (R²=0.658, p<2.2×10^-14) of the RBFN model.

Key words: RBF neural network, urban air quality, causal factors

摘要：

基于2010年我国78个主要城市的5个空气质量指标数据, 利用主成分分析方法, 计算城市空气质量综合得分; 采用多元线性回归方法, 从气候、地形、城市发展和城市环境状况四方面的48个变量中筛选出与城市空气质量显著相关的10个, 并定量评价不同因子对城市空气质量的贡献。依据空气质量综合得分, 从78个城市中分层随机选取30个城市作为训练样本, 构建基于径向基函数(RBF)的人工神经网络模型。利用城市自然、社会、经济特征及环境管理现状模拟城市空气质量, 并应用于我国173个主要城市空气质量状况的模拟。结果表明, 年平均饱和水气压、城市建成区面积、城区海拔落差和第二产业占GDP的百分比是影响中国当前城市空气综合质量的主要因素, 分别可以解释城市空气质量变异性的14.7%, 12.8%, 8.8%和7.2%。研究结果突破了以往大部分空气质量评价模型仅以大气污染物和气象因子作为模型输入因子的局限, RBF人工神经网络模型的模拟结果准确性较高(R²=0.658, p<2.2×10^-14)。

关键词: 人工神经网络, 城市空气质量, 影响因子