評估台灣回溯式PM2.5之預測情形：模式間之結果比較

Abstract

近幾年許多流行病學研究顯示，長期暴露於大氣中空氣污染物細微顆粒物PM2.5會使人體健康產生肺癌以及心血管功能傷害相關的疾病，甚至造成死亡率的攀升。然而，早期PM2.5的測量機制並不是那麼完整而限制了分析長期PM2.5暴露對健康造成的影響。本研究以相關空氣污染物和氣象條件資料發展回推預測PM2.5濃度之模式。 從2005年台灣環境保護署空氣品質監測網普通測站開始有完整的PM2.5量測數據。為預測過去缺少之PM2.5測量值數據，將現有的空氣品質監測站和中央氣象局測站兩個單位資料分為2005-2009年和1993-2004年兩組數據分別用於建立和驗證模式。運用每日平均濃度的一氧化碳，氮氧化物，二氧化硫，和臭氧空氣污染物以及溫度，風速和相對濕度等氣象條件，建立與PM2.5/PM10比值相關的各類混合模式。將回推預測與實際量測之PM2.5濃度值作迴歸模式，比較R2值來驗證模式之可信度，並應用推估結果來評估比較模式的適用性。 以PM2.5的回推結果而言，針對全國空氣品質監測站，運用2005-2009年數據建立模式所預測PM2.5的數據與實際測量值可顯示良好的相關性(R2=0.62－0.92)，除了台東測站有偏低之相關性(R2=0.32－0.40)。針對從1997年開始有PM2.5量測數據的古亭，忠明，鳳山和林園四站空品測站，比較所回推預測的PM2.5與1997-2004年實際測量PM2.5的數據也呈現良好的相關性(R2=0.66－0.72)。另外，除了氣狀空氣污染物，只運用PM10，能見度，溫度，風速和相對濕度資料納入模式分析預測和實際量測的PM2.5就能指出良好的相關性(R2=0.66－0.71)。 根據空氣污染物，能見度和氣象變項條件可運用PM2.5/PM10比值預測過去PM2.5之暴露情形。但是，資料來源的監測數值資料不足的情形會限制模式的建立和預測回推值所建立的模式，以致於影響推估結果；而污染物濃度在不同空間尺度上的變化也會造成不同的推估結果，因此未來可依據各測站的資料筆數完整性和測站地理位置以及特性來分別建立不同之模式。

Many epidemiological studies have found increased mortality as well as disease, such as lung cancer and cardiopulmonary morbidity were associated with long-term exposure to ambient fine particulate matters, PM2.5 (particulate matters with aerodynamic diameter ≦2.5 μm). However, ambient monitoring programs for PM2.5 did not exist in early years, posing difficulties on analyzing health effects from long-term PM2.5 exposure. This thesis developed site-specific models for retrospective prediction of PM2.5 levels using available data on air pollutants and meteorological variables. The Taiwan Environmental Protection Administration has complete PM2.5 data at each monitoring site since 2005. In order to predict values during periods when PM2.5 data were not available, the dataset from air quality monitoring and central weather bureau stations were divided into two groups with data in 2005-2009 and in 1993-2004 used for model building and verification, respectively. The data were used to model PM2.5/PM10 ratio with daily 24-hour average levels of CO, NOx, SO2, and/or O3. In addition to air pollutants, visibility and meteorological variables including daily temperature, wind speed, and relative humidity were also considered. Models were developed separately for each air monitoring site. PM2.5 estimation results for each air monitoring site from 2005 to 2009 matched well with the actual PM2.5 data (R2=0.62-0.92), except for the Taitung site (R2=0.32-0.40). For Guting, Chungming, Fengshan and Linyuan sites which had complete PM2.5 data from 1997, comparison of the PM2.5 estimation for 1997-2004 with the measured PM2.5 also shows moderate association (R2=0.66-0.72). With only PM10, visibility and the meteorological data, but not gaseous pollutants, included in the analysis shows strong association (R2=0.66-0.71). Our study results show that it is feasible to use PM2.5/PM10 ratio to predict historical PM2.5 exposure levels from existing data of air pollutants, visibility and meteorological variables. The variability of pollution concentration in different spatial scales could affect the modeling results. Thus, establishing empirical models separately for different types of monitoring sites may be necessary.