Statistical modeling of air pollution index based on its structure and descriptive healthy status

Abstract

Accurate air pollution modeling is essential for estimating the Air Pollution Index (API) effectively. It‘s because, the air quality assessment depends on the ability of selecting the suitable probability density function (PDF) to describe the observed air pollution data. This thesis characterizes the API hourly data of several cities in Klang Valley region, Malaysia, for the period of January 2005 to December 2014. Three different statistical approaches are proposed for modeling API data, including conventional models, API structure models, and descriptive status models. The first approach is the conventional models, which considers several common distributions used for modeling the API and its related pollutants. Several selection criteria are employed to select the best fitted distribution and the weight of ranks method is proposed to solve the confliction between criteria in selecting the best model. The fitted distributions of the observed and generated API data are utilized for comparisons to other proposed models. In addition, the selected conventional distributions are used as a basis in the construction of API structure models. The second approach is the API structure models, which involve a mixture of distributions for the critical pollutant variables. Furthermore, the third approach was based on the descriptive status of the API. The results show that the healthy status is able to be described using the conventional fitted models, while the Generalized Pareto Distribution (GPD) is found to be a good fitted model for the unhealthy status. Indeed, based on the selection criteria, it was found that the API structure models are superior for modeling the API data while the API descriptive status models are useful for evaluating the unhealthy API return level. Next, the estimation of unhealthy API recurrence events is improved using a Markov chain model based on Generalized Pareto Distribution (MC-GPD) and Hierarchical-Generalized Pareto model (HM-GPD). The MC-GPD is used to model the serial dependence consecutive values of unhealthy API in the hope of capturing and exploiting as much information correspond to the threshold of unhealthy environment. In MC-GPD model, the serial dependence has been modeled by using the first order Markov Chains. The parameters for a single GPD model are combined together with series dependency parameters for the transformation of marginal model for each station to the Frechet unit using the likelihood of the Markov chain model. The Hierarchical-Generalized Pareto model is utilized to integrate the information about location and seasonal effects from the marginal GPD models of hourly API exceedance data, along with the information of the serial dependence at each location. The MC-GDP and HM-GPD models are found to be able to provide a precise estimation of the unhealthy API return level. The accuracy of return level estimates for unhealthy API at single sites are found to be improved by employing the MC-GPD model which that takes advantage of the climatological dependence structure in the API data. Overall, we conclude that the mixture distribution of the API components should be considered as a better method for estimating the API data. Apart from that, we conclude that the parameters estimated form the MC-GPD is found to be able to provide a better estimate of the return levels of unhealthy API.