Estimation of Near Ground Particulate Matter in Urban Areas
MetadataShow full metadata
Air quality and its effect on human health is an area of increased research and interest over the last twenty years. As the world’s population increases understanding the effects of human activity on the environment and air quality becomes even more important. The health effects of PM10, the default priority pollutant for New Zealand, was quantified by Kuschel, et al. (2012). They estimated that in 2006, over 600 premature deaths in Auckland were related to air quality. They also reported a cost of over $2 billion to the Auckland region in 2006 as a result of exposure to ambient PM10. Despite this high economic and social cost studies in the literature, both grey and white, related to prediction of Auckland’s PM10 are sparse. Most of the PM10 models in the literature are highly dependent on the input data used. Each model uses different inputs making it hard to compare the effectiveness and evaluate the generalisability of these models. The data used is largely opportunistic – use what we have – rather than informed. Moreover, for many regions including Auckland access to data such as detailed emission inventories, land use, and demographic distributions is not always possible. Hence, the methods in the literature have a limited practical use. This thesis aims to answer questions related to Auckland’s site-specific PM10 concentrations, including PM10 trends, relative contribution of meteorological sources, and one day ahead prediction of PM10 concentration. Semi-empirical, statistical, and geo-statistical methods are explored. Attempts to tackle the challenges of modeling a nonlinear system by using Artificial Neural Networks (ANNs), Long short-term memory (LSTM), and Random Forest (RF) methods are reported. These models are parsimonious and make use of routinely available meteorological data collected from the six fully operational monitoring stations in Auckland during 2011-2016. It was found that Auckland’s PM10 has complex seasonal patterns and that PM10 concentration trends are very localized and cannot be fully explain by land usage (rural vs urban). Using GAM and GAMM models, not previously used for Auckland, a clear difference was found between the effects of temporal aspects of anthropogenic sources and atmospheric conditions on PM10. For modelling with linear statistics, the main challenge encountered was to find a way to characterise the spatio-temporal dependence structure. The inability to accurately and fully define this structure limits the usefulness of these linear approaches. In contrast it was found that machine learning (MLP, LSTM) and ensemble methods RF were able to account for this underlying structure and for the dynamism of the process. Of all the methods explored the RF model was found to be the most accurate and therefore the most promising avenue for future work.