中国有效实施了一系列大气污染治理措施，细颗粒物（PM2.5）的污染浓度有所降低。但是，目前中国空气质量和世界发达国家水平仍有较大差距，人群健康风险依然不容忽视。为制订科学有效的污染防治策略，实现PM2.5污染的持续改善，亟需进行精细化PM2.5来源解析，支撑控制路径的成本效益分析工作。PM2.5浓度同前体物排放间的非线性响应关系十分显著，传统空气质量模拟对计算资源需求相对较高，为PM2.5 源解析和污染防治策略的效果评估带来了挑战。现有研究通过行业线性拟合和微分方法搭建了曲面响应模型（RSM-SL (DM)），并在污染物源解析和政策效果评估中予以应用。然而，建立RSM仍需要在前期进行大量的空气质量模拟；倘若减少模拟次数，RSM的准确性便可能下降。近期研究基于神经网络技术（NN-CTM）实现了对CMAQ模拟结果（CMAQ-CTM）的较好再现，其对中国2015年四个典型月份PM2.5浓度模拟的平均绝对误差（MAE）仅有1.46 μg/m3。在上述研究基础之上，本研究扩充了建模所需的输入样本，建立了高分辨率的NN-CTM RSM方法。NN-CTM RSM可服务于PM2.5源解析和污染治理措施的成本效益分析，并同大气污染防治综合科学决策支持平台（ABaCAS）耦合而得到最优的PM2.5 污染控制策略。研究也将NN-CTM RSM与其他方法进行对比，并详细分析了NN-CTM RSM的优势。本研究优化建立的NN-CTM RSM-v2相对传统神经网络方法能够提升对PM2.5、SO2和NO2的模拟准确性。基于建模结果，研究也进一步指出了现有机器学习模型用于空气质量模拟的挑战。例如，受前体物排放与浓度分布的非平稳性以及气象等外部因素的综合影响；空气质量模拟的较高计算成本也限制了可用样本的数量；可用样本数量还会对模型的复杂性和训练时长产生影响，而充足的训练时长对不可见样本的准确预测至关重要。本研究对上述模型调优的具体难点展开了详细分析，完善了机器学习方法应用于空气质量模拟的理论基础，提出了神经网络方法用于空气质量高分辨率模拟的应用方案。
PM2.5 (particulate matter less than or equal to 2.5 μm in diameter) pollution is a significant contributor to worldwide mortality, and its reduction is a key target of air pollution control initiatives. Policy makers seek source apportionments and control policy cost-benefit analyses to aid in the development of effective emission control strategies. The highly non-linear response of PM2.5 concentration to emissions of its precursors presents challenges to many source apportionment methods, and the computational cost of Air Quality Model (AQM) simulations present a barrier to use of AQM-based techniques in policy applications. Response surface modeling (RSM) coupled with sectoral linear fitting and differential method (RSM-SL) has been successfully used to perform source contribution and control policy analysis. However, building the RSM still requires many high cost AQM simulations. A neural net based CTM approximation (NN-CTM) has recently been trained to emulate the CMAQ-CTM with MAE of 1.46 μg/m3 in predicted PM2.5 concentration over China for four months in 2015.In this work, we attempted to produce a neural network robust enough to produce a “high resolution” NN-CTM RSM by using many more samples than AQM based RSMs. In theory, the NN-CTM RSM could then be used to perform source apportionment, perform cost-benefit analyses of PM pollution reduction, and derive optimal control strategies in combination with the Air Benefit and Cost and Attainment Assessment System (ABaCAS). This work has comprehensively compared the prediction accuracy of the new NN-CTM RSM architecture with other modeling methods, and identified the overall improvements in predicting PM2.5, SO2 and NO2 concentrations. Based on the modeling experiments and analysis, this work also points out a number of significant barriers to tuning a neural network to the degree of accuracy and robustness necessary to use it in downstream applications, such as: the nonstationary distributions of concentrations of precursors, emissions, and external factors such as meteorological conditions; that the training of the neural network still requires samples generated from an AQM; that the number of available samples limits the complexity of the neural network to be trained; finally, that the AQMs which we seek to emulate with our neural network are themselves inaccurate. This work discusses these barriers in detail, establish best practices for future attempts to incorporate machine learning into air quality modeling, and lay the theoretical groundwork for future extensions such as the NN-CTM RSM proposed above.
