Study On The Forecast On Surface Ozone Exceedence Days Based On Non-Gaussian Distribution

Since prolonged exposure to high surface ozone environment was hazardous to humans. For the sake of human health, it is very important to forecast harmful surface ozone concentration with satisfactory accuracy. In this thesis, pollutants and meteorological data monitored in San Francisco Bay Area and Houston Metropolitan are selected to build prediction models. Due to the non-Gaussian distribution of surface ozone data, the non-parametric T2control limit in principal component subspace method and a generalized linear mixed model (GLMM) are proposed to forecast surface ozone exceedence days. The main works and results include:(1) The non-parametric T2control limit in principal component subspace method is used to forecast ozone exceedence days. Since the huge difference between the hypotheses of Gaussian distribution in Hotelling T2control limit and the real distribution of measurement data, some surface O3exceedence days cannot be predicted. A non-parametric T2control limit of principal component subspace based on the real distribution of measurements data is proposed to forecast ozone exceedence days.(2) GLMM model is also proposed to predict surface ozone concentration. Due to the different impact of meteorological factors in different synoptic patterns on surface ozone build-ups and the non-Gaussian distributed surface ozone concentration, high-level surface ozone concentration cannot be predicted correctly by conventional linear regression model. Thus, a cluster analysis is used to divide modeling data into different synoptic patterns. Then, a generalized linear mixed model (GLMM) is employed to predict surface ozone concentration. Because of the difference surface ozone build ups and the characteristics of surface ozone distribution are considered by GLMM, it outperforms conventional linear regression model in predicting high-level surface ozone concentration.(3) The performance of two models is compared. It is shown that all exceedence days in Livermore Valley can be correctly predicted by PCA based model, and5of11ozone exceedence days can be forecasted by GLMM model; in Deer Park site, two surface ozone exceedence days are removed when impute the missing data by GLMM, and during the prediction process, PCA based model have two more exceedence days than GLMM.5of8ozone exceedence days can be forecasted by PCA based model, and5of6can be forecasted by GLMM. Although, the characteristic of non-Gaussian distribution of surface ozone concentration is considered by both models, the non-parametric T2control limit is determined by the real distribution of measurement data. PCA based model outperforms GLMM when modeling unimodal distributed surface ozone data in Livermore. GLMM gives a better prediction when modeling bimodal distributed surface ozone data in Deer Park, because bimodal distributed surface ozone data are considered as two unimodal distributions by GLMM, while surface ozone data were not classified in PCA based model. Therefore, unimodal distributed surface ozone exceedence days can be forecasted by PCA based model, while GLMM has a better performance on predicting bimodal distributed surface ozone concentration.