【作者】 张强；

【导师】 杨光飞；

【作者基本信息】 大连理工大学 ， 环境工程， 2023， 硕士

【摘要】 随着经济的快速发展,全球工业化以及城市化进程的加快,PM_2.5等细颗粒物污染已经成为危害人们身心健康,阻挠社会发展的重大环境问题。中国针对PM_2.5的治理已经进入深水区,需要探寻精确的PM_2.5时空分布模式和污染传播机制,形成PM_2.5污染长期预报预警机制及跨区域协同治理、联防联控制度体系。然而,由于PM_2.5分布具有时空异质性、溢出效应,现有文献中鲜有针对长时间尺度下污染传播模式的研究,且现有研究中的预测手段无法进行长期精确的预测,导致目前PM_2.5治理陷入瓶颈。基于此,为实现对污染溢出效应的精确分析及PM_2.5长期精确预测,本文首先利用PM_2.5污染数据,使用离散小波变换对原始PM_2.5序列进行分解,在每个时间尺度上,使用BEKK-GARCH模型验证并构建各个城市之间的PM_2.5污染溢出关系,并基于复杂网络理论,构建了每个时间尺度下的城市间的PM_2.5污染溢出网络,针对污染溢出网络进行了多视角分析。其次,在溢出效应的基础上,结合气象数据,借助深度学习模型和经验模态分解算法,成功构建了专门针对长时空序列预测的深度学习模型GAT-EGRU。本研究的主要研究结论如下:（1）PM_2.5污染在空间上呈现出关联性,城市之间存在明显的污染溢出关系;不同时间尺度上中国城市的污染溢出网络拓扑结构特征表现出时间异质性。网络拓扑结构揭示了污染溢出网络随着时间推移呈现明显的演化特征,污染发生后的48～96小时内污染溢出效应最为显著;在污染溢出网络的个体特征方面,度中心性排名靠前的城市在网络中占据主导地位,介数中心性排名靠前的城市在网络中扮演桥梁的角色,紧密中心性较高的城市更容易受到污染溢出效应的影响,特征向量中心性分析揭示了南方大污染集群;社团划分结果显示随着时间尺度的推移,南北两个大污染集群不断扩张,对研究区域的污染溢出产生影响。（2）针对长时间尺度的城市PM_2.5预测,本研究结合经验模态分解算法构建了深度学习模型GAT-EGRU。为验证模型有效性,进行了模型对比、真实值对照和蒸馏性实验。结果表明,在6个时间尺度的12次比较中,该模型获胜了11次。模型在48小时内的预测结果的均方根误差（RMSE）为21.65,平均绝对误差（MAE）为15.14,在长时间尺度上明显优于同类模型;与真实值的对照实验证实了该模型的有效性。蒸馏性实验验证了模型各个模块的作用,并解释了模型在预测准确度和计算复杂度方面的权衡。本研究成功建立了以城市为主体的PM_2.5长时空分析预测框架,丰富了现有PM_2.5长时空分析预测建模的方法体系,有助于政府制定合理的污染管控政策及策略。更多还原

【Abstract】 With the rapid development of the economy and the acceleration of global industrialization and urbanization,the pollution of fine particulate matter such as PM_2.5 has become a major environmental problem that harms people’s physical and mental health and impedes social development.China’s governance of PM_2.5 has entered a deep water zone,requiring the exploration of accurate PM_2.5 spatiotemporal distribution patterns and pollution propagation mechanisms,forming a long-term forecasting and warning mechanism for PM_2.5 pollution,and a cross-regional collaborative governance and joint prevention and control system.However,due to the spatiotemporal heterogeneity and overflow effects of PM_2.5 distribution,there is a lack of research on pollution propagation patterns at long-time scales in existing literature,and the existing prediction methods cannot make long-term accurate forecasts,resulting in the current bottleneck of PM_2.5 governance.Based on this,this article uses PM_2.5 pollution data and uses discrete wavelet transform to perform multi-scale decomposition on the original PM_2.5 spatiotemporal sequence.Then,at each time scale,the BEKK-GARCH model is used to verify and construct the PM_2.5 pollution spillover relationship between cities.Next,based on complex network theory,a PM_2.5 pollution spillover network between cities is constructed at each time scale,and a multi-perspective analysis of the pollution spillover network is conducted.Secondly,on the basis of the spillover effect,combined with meteorological data,and using deep learning models and empirical mode decomposition algorithm,the GAT-EGRU deep learning model specifically for long spatiotemporal sequence prediction was successfully constructed.The main conclusions of this article are as follows:（1）Regarding the spatial correlation of urban PM_2.5,there is a clear correlation in PM_2.5 pollution in space,and there is a significant spillover effect of PM_2.5 pollution among cities.In terms of the topological structure of the pollution spillover network at each time scale,the pollution spillover network of Chinese cities has significant temporal heterogeneity.The overall topological structure of the network shows that the pollution spillover network has a significant evolution process as the time scale advances,and the pollution spillover effect is most significant 48 to 96 hours after pollution occurs.Regarding the individual characteristics of the pollution spillover network,cities with high degree centrality rank dominate the pollution spillover network,cities with high betweenness centrality rank act as bridges in the pollution spillover network,cities with high closeness centrality are more susceptible to the pollution spillover effect,and cities with high eigenvector centrality reveal the existence of the southern major pollution cluster.The results of pollution community partitioning show that the two major pollution clusters in the north and south expand with the advancement of the time scale,affecting the pollution spillover in the study area.（2）For the prediction of long-step urban PM_2.5,a deep learning model GAT-EGRU was successfully constructed by combining empirical mode decomposition algorithm.To verify the effectiveness of the model,comparative experiments,real value contrast experiments,and distillation experiments were conducted.The experimental results demonstrate that the model won 11 times out of 12 comparisons in six time scales,outperforming similar models and far superior to similar models in long time scales.The real value contrast experiment verifies the effectiveness of the model.The distillation experiment verifies the function of each module of the model and explains the balance between prediction accuracy and computational complexity.This study has successfully established a long-term and spatial analysis and prediction framework of PM_2.5 with cities as the main body,which enriches the method system of PM_2.5long-term and spatial analysis and prediction modeling.This framework can help the government to formulate reasonable pollution control policies and strategies.更多还原