Understanding and Attributing Extreme Air Pollution Events in a Changing Climate

空气质量差是影响所有工业化国家和整个社会的全球健康问题。它的特点是颗粒物、臭氧和氮氧化物的有害水平，慢性和急性接触这些物质可能会增加发病率。全世界每年约有300万人死于空气污染。在英国，每年有数万人过早死亡，许多地方的空气质量超标。本博士的目标是开发一系列基于测量空气污染物浓度的统计模型，以：1)表征极端空气污染事件的频率、持续时间和空间范围；2)调查极端污染事件的气象驱动因素；2)评估气候变化如何影响极端事件的发生。该项目将分三期进行。1. 表征英国极端空气质量事件英国的AURN网络由约200个测量点组成，用于确定空气污染物的浓度和趋势。每小时的AURN数据用于警告公众空气质量不良，并报告是否符合欧盟空气质量指令。该项目的第一步是(A)熟悉AURN网络，包括污染物（PM, O3, NO2）的行为和在不同地点类型（例如城市/农村/路边）观察到的趋势，以及(b)编写代码读取/处理数据。利用具有较长记录（bbb20年）的站点，对极端空气污染事件进行综合评价。利用极值分析，将开发统计模型来估计污染物超过空气质量指导阈值的可能性。将调查极端事件的大小和可能性的地点和区域变化。极端污染事件的气象驱动因素为了准确预测极端污染事件并量化这些事件的特征是否以及如何变化，了解极端污染事件本身与任何关键气象驱动因素之间的联系至关重要。我们还想了解在极端和非极端情况下，污染物及其驱动因素之间的关系有何不同。这些问题将使用极值协变量建模方法来解决，模型最初在现场的基础上进行拟合。我们将研究驱动-响应关系的参数和半参数协变量模型，通过使用滞后的驱动变量和/或直接对响应中的序列极值依赖性进行建模，尽可能构建可用于短提前期预测的模型。3.开发英国空气污染的时空模型极端空气污染事件的长期趋势或周期的物理驱动因素存在不确定性。为了补充(2)中的工作，我们将使用极值潜在过程模型，其中使用潜在过程将特定地点的参数建模为时间和/或空间上的函数，以产生这些过程行为的数据驱动可视化。与基于回归的模型相比，这种方法的优点是：(i)不需要了解驱动因素的随机行为，并且对驱动因素的随机行为进行最小的先验假设；（ii）它们自然地扩展到包括时间和/或空间依赖结构。包括依赖性对于提高对物理过程（污染物和驱动因素）的科学理解以及从统计模型中获得精确和有效的参数和不确定性估计都是重要的。关于极端事件的潜在过程建模的使用有相当大的文献，主要是通过贝叶斯分层建模框架的实现。动态线性模型，在极值上下文中研究得很少，也可能是捕获时间变化的有用框架。这个博士的新颖性将是这些模型类型在英国空气污染的应用和发展

Poor air quality is a global health concern affecting all industrialised nations and all society. It is characterised by harmful levels of particulate matter, ozone and nitrogen oxides for which chronic and acute exposure to may increase morbidity. Worldwide, ~3 million deaths are attributable to air pollution annually. In the UK, tens of thousands of premature deaths occur annually and many locations breach air quality guidance thresholds. The objective of this PhD is to develop a series of statistical models based on measured air pollutant concentrations to: 1) characterise the frequency, duration and spatial extent of extreme air pollution events; 2) investigate meteorological drivers of extreme pollution events and; 2) assess how climate change may influence the occurrence of extremes. The project will be conducted in 3 phases. 1. Characterising UK extreme air quality episodesThe UK's AURN network consists of ~200 measurement sites and is used to determine the concentration and trends of air pollutants. Hourly AURN data are used to warn the public of adverse air quality and for compliance reporting against EU Air Quality Directives. A first step in this project will be to (a) gain familiarly with the AURN network, including pollutant (PM, O3, NO2) behaviour and observed trends at different site types (e.g. urban/rural/roadside), and (b) to write code to read/process the data. A comprehensive evaluation of extreme air pollution events will be performed, exploiting sites with long records (>20 years). Using Extreme Value Analysis, statistical models will be developed to estimate the likelihood of pollutants exceeding air quality guidance thresholds. Site-wise and regional variability in the magnitude and likelihood of extremes will be investigated.2. Meteorological drivers of extreme pollution episodesTo both accurately forecast extreme pollution events and quantify if and how the characteristics of these are changing, it is vital to understand the connection between the episodes themselves and any key meteorological drivers. We would also like to understand how the relationship between pollutants and their drivers differs between extreme and non-extreme scenarios. These questions will be addressed using extreme value covariate modelling methods, with models fitted initially on a site-wise basis. We will investigate parametric and semi-parametric covariate models for the driver-response relationship, trying where possible to construct models that can be used for short lead-time forecasting, by using lagged driver variables and/or directly modelling the serial extremal dependence in the responses. 3.Developing spatiotemporal models for UK air pollutionThere is uncertainty in the physical drivers of long-term trends or cycles in extreme air pollution events. To complement work in (2), we will use extreme value latent process models, in which site-specific parameters are modelled as functions in time and/or space using latent processes to produce data-driven visualisations of the behaviour of such processes. The advantages of this approach over the regression-based model are that (i) no knowledge of, and minimal prior assumptions on the stochastic behaviour of, the drivers is needed and (ii) they naturally extend to include temporal and/or spatial dependence structures. Inclusion of dependence is important both in improving scientific understanding of the physical processes (both pollutants and drivers) and in obtaining precise and efficient parameter and uncertainty estimates from the statistical model. There is a relatively large literature on use of latent process modelling for extreme events, primarily through implementation of a Bayesian Hierarchical modelling framework. Dynamic linear models, much less well explored in an extreme value context, may also be a useful framework for capturing temporal changes. This PhD's novelty will be in the application and development of these model types for UK air pollu