基于多源时空数据的雾霾重污染过程动态特征分析与预警研究

Both the underestimation of heavy haze peak values and the delay of recognition to the variation tendency generated by the current early warning system for haze will result in false positive or false negative of heavy haze episodes, which results from absence of the key information including distribution of precursors, regional transport and transition of the external environment. In this study, dynamic spatial-temporal features of the processes for heavy haze pollution will be focused on; and the emphasis is the accurate simulation of the peak values and recognition of the inflexion point for heavy haze episodes. Multiscale convolutional long short-term memory (MC-LSTM) artificial neural network will be employed to analyze the surface and high altitude meteorological data, air pollutant data, traffic data and the change of emissions due to the implementation of emergency actions during the haze episodes. Environmental data from multi-sources are used to infer the real-time precursor emission; and the synoptic patterns in the studied area are used to describe the strength of regional transport. The processes of heavy haze episodes will be divided into several stages according to the change of interactions among the input variables. The global temporal-spatial dynamic characteristics extracted from environmentally big data by the deep neural network will be used to recognize the inflexion point of heavy haze pollution. The hidden features extracted by the deep neural networks shall be explained by the decision tree in the semantic level and then the dominant factors of different stages for heavy haze pollution can be explained quantitatively. Weights of the dominate variations will be strengthened in different stages according to the mechanism knowledge of environmental chemistry, which is supposed to reduce the simulation error of the peak value. Artificial intelligence and human intelligence will be integrated to develop the hybrid model driven by data and mechanism. It can not only improve the success index of heavy haze pollution prediction, but also provide scientific basis for the utilization of environmentally big data from multi-sources intellectively.

由于对包含前体物分布、区域输送和外部环境系统性转换等关键信息的多源数据利用率较低，现有预警模型对雾霾重污染峰值模拟偏低，演化趋势识别滞后，时常出现漏报和误报。因此，项目拟集中研究雾霾重污染的动态时空特征，重点实现其峰值的精确模拟和消散拐点的及时识别：采用多尺度卷积-长短记忆神经网络分析重污染期间污染物、交通流量和产业分布等数据，推断实时污染排放情况，结合中尺度气象场描述区域输送强度；通过研究各要素之间作用关系的变化将雾霾重污染分为若干阶段，使用深度神经网络获得多源数据中的全局时空信息特征，实现消散拐点的准确识别；运用决策树方法计算卷积核对最终输出结果的贡献率，有效识别各阶段影响雾霾浓度的主导变量；利用环境化学机理知识强化主导变量权重，降低峰值模拟误差。通过在人工智能框架上融合人类智能建立机理-数据混合驱动模型，不仅可有效解决雾霾重污染预警问题，还可为“智能化”利用多源环境数据提供理论依据。

由于对包含前体物分布、区域输送和外部环境系统性转换等关键信息的多源数据利用率较低，现有预警模型对雾霾重污染峰值模拟偏低，演化趋势识别滞后，时常出现漏报和误报。本项目在保证污染事件样本容量的基础上定义重污染数据样本，利用华北平原空气污染数据和气象数据建立重污染事件数据库，并采用有约束动态时间规整算法解决时间不等长问题；通过多项主元分析模型实现污染事件演化阶段自动划分，并通过主元空间贡献率识别各阶段主导变量识别，发现产业分布、地理位置和地形分布对污染主导变量具有重要影响。在此基础上，针对LSTM可解释性较差、监测数据的多源异构特点和雾霾重污染时间序列多尺度、非平稳特征开发了基于深度时间序列特征融合的雾霾重污染过程预警，对雾霾严重污染样本的准确预测率分别为94.12%、85.29%、77.57%和51.10%，显著高于已有应用模型；当前雾霾浓度对未来浓度的影响随预测步长的增加从80.89%（t + 3）急剧降低至16.34%（t + 24），前体物浓度影响力从5.23%（t + 3）上升至29.43%（t + 24）。为进一步提高雾霾浓度预警精度，采用prophet算法获取雾霾浓度的变化趋势、日周期、周周期和年周期并进行LSTM建模，6小时预测的R2达到0.85，MAE为15.41 μg‧m-3，对重污染样本的预测准确率为85.29%。为提取地表监测数据的空间特征，建立了2DCNN-LSTM模型针对成都盆地雾霾浓度进行预测，通过皮尔森相关系数和HYSPLIT选择空间站点，加入CNN结构MAE降低26.10%，当使用PCA压缩模型参数后，拟合优度提高5.50%。使用Prophet提取各个省会城市的PM2.5污染物趋势特征并使用硬聚类方法划分空气质量地域特征，使用向量误差修正模型对部分典型省会城市的雾霾趋势、气象、风速、工业企业数量、工业烟尘排放量、二氧化硫排放量和货运总量进行建模，获得空气质量改善的主要贡献因素：工业烟尘排放量是郑州、石家庄、成都和北京的雾霾方差主要贡献变量；工业企业数量影响沈阳和武汉的雾霾浓度；平均风速变化影响广州和杭州的雾霾浓度；呼和浩特和西安主要受平均温度影响，结果基本符合中国经济发展的整体布局。