中文摘要：

针对PM2.5日均质量浓度，采用BP人工神经网络模型，预测研究区空气中PM2.5浓度的空间变异，通过与普通克里格（OrdinaryKriging）插值方法对比验证BP人工神经网络预测模型的精度。结果表明：BP人工神经网络预测模型下研究区检验样本点位置的PMz5仿真浓度与观测浓度之间的均方差、平均绝对误差、平均相对偏差和相关系数分别为0．296μg0／m0、0．412μg／m3、1．650％和0．851；而与此同时，普通克里格插值方法下的对应结果分别为1．041μg0／m0、0．689μg／m3、11．910％、0．638。研究成果在肯定BP人工神经网络预测模型可用于揭示PMz．5浓度空间变异特征的同时，也证实了其相对于普通克里格插值方法在固定空间点位准确预测PM2s浓度方面的优势。

英文摘要：

Using back-propagation （BP） artificial neural network model, this study attempted to predict the spatial variability of PM2.5 daily average concentrations in an urban area with the results from Ordinary Kriging （OK） as a bench mark. Results shows that while the mean square error （MSE）, mean absolute error （MAE）, bean relative deviation （A） and correlation coefficient （R2） between BP simulated PM2.5 daily average concentrations and observed ones are 0. 296 /μgZ/m6 , 0. 412 μg/ma , 1. 650 % and 0. 851, respectively, those between OK-based predicted PM2.5 daily average concentrations and observed ones are 1. 041 μgZ/ms ,0. 689 μg/m3 ,11. 910% and 0. 638. It can be concluded that BP artificial neural network prediction model is a promising approach in revealing spatial variability characteristics of urban PM2.5 pollution, and can pro-vide simulated PM2.5 concentration with greater accuracy compared to OK method.