The Distribution of PM2.5 in Kaohsiung City during 2012-2016 and the Establishment of PM2.5 Prediction Models

本研究彙整2012-2016年高雄市一般空氣品質監測站資料，分析PM2.5濃度分布及其變化，結果顯示冬季為PM2.5污染及能見度問題最為嚴重之季節，夏季污染及其問題則最為輕微。以PM2.5日均值15、35、50及70 μg/m3作為標準，劃分為五個等級之濃度區間，結果顯示各測站事件日（超過35 μg/m3）日數有逐年減少的趨勢，且事件日逐漸以35～50 μg/m3為主，2016年該等級日數佔總事件日日數之比例達60.3～79.5%。探討各測站冬季時，氣象因子對PM2.5分布的影響，並比較事件日發生的機率。結果顯示不同氣溫下，各測站事件日發生的機率差異小。各測站於低相對濕度下或無雨時發生事件日的機率較高。除小港及林園測站以外，其他測站於低風速下事件日機率較高。另外，除了林園外，其他測站PM2.5事件日的發生與靜風時數有關。無降雨的條件下，PM2.5污染有逐年改善趨勢，能見度問題則未獲得明顯改善。對各測站分季建立複線性迴歸模型，以預測次日PM2.5日均值，利用2016年資料作為測試資料，驗證模型性能，結果顯示春季模型預測性能最佳，R2值達0.65～0.80；就預測濃度區間而言，夏季模型的預測準確率最高，達98%以上，次高為春、秋季的模型，準確率分別達71.1～83.1%及67.1～89.9%，最低為冬季，準確率僅達50.0～64.4%。In this study, the data were collected from general air quality monitoring stations in Kaohsiung City during the period from 2012 to 2016, and been analyzed for the changing of PM2.5 concentration distribution. The pollution of PM2.5 and visibility problem are the most serious in winter, and it are the most trifling in summer. Based on the daily average of 15, 35, 50 and 70 μg/m3, divided into five levels. The number of PM2.5 pollution episodes at each stations have decreased year by year, and the main of the daily average is about 35～50 μg/m3 gradually, the ratio of the episodes to the total episodes was 60.3～79.5% in 2016. This study investigated into the influence of meteorological factors on the distribution of PM2.5, and the difference of probability of the episode. The results show that the difference of probability of the episode is small at different temperatures, and the higher probability of occurrence of the episode at low relative humidity or no rain at each stations. Besides Siaogang and Linyuan, the higher probability of occurrence of the episode at low wind speed at each stations. In addition, the occurrence of the episode in relation to the hours of calm winds except Linyuan. Under the condition of no rainfall, the pollution of PM2.5 has been improving year by year, but the problem of visibility has not been improved significantly. This study established the multiple regression model for predicting the next day average PM2.5 concentration, and using data in 2016 as testing data to verify model performance. The results show that the performance of the spring models are the best, R2＝0.65～0.80. In terms of the predicted concentration interval, the accuracy of the summer models are the highest, above 98%. The predictive performances of the spring and the autumn models are also good, 71.1～83.1% and 67.1～89.9% respectively. However, the accuracy of the winter models are the lowest, only 50.0～64.4%.摘要 i Abstract ii 目錄 iii 表目錄 v 圖目錄 vi 第一章 前言 1 1.1. 研究緣起 1 1.2. 研究目的 2 第二章 文獻回顧 3 2.1. 高雄地理環境與污染源分布 3 2.2. PM2.5 形成機制與化學組成 6 2.3. PM2.5的危害 8 2.4. PM2.5與氣象因子的關係 9 2.5. PM2.5濃度預測之模型 10 第三章 研究方法 13 3.1. 研究架構 13 3.2. 監測站位置及監測項目 14 3.3. 資料前處理 17 3.4. 相關性分析 21 3.5. 迴歸分析 21 第四章 結果與討論 24 4.1. 高雄市PM2.5濃度及事件日之分布 24 4.1.1. PM2.5小時值分布 24 4.1.2. PM2.5事件日分布 32 4.2. PM2.5與氣象因子之關係 36 4.2.1. 氣溫 36 4.2.2. 相對濕度 40 4.2.3. 風速 44 4.2.4. 降雨 52 4.3. 能見度與PM2.5及RH之關係 55 4.4. PM2.5預測模型 68 4.4.1. 相關性分析 68 4.4.2. 複迴歸模型之建構過程 73 4.4.3. 迴歸模型預測性能之驗證 81 第五章 結論與建議 92 5.1. 結論 92 5.2. 建議 94 參考文獻 96 附錄A 資料遺失值數量及比例 100 附錄B PM2.5小時值分布 109 附錄C 各等級事件日季節分布 118 附錄D 各模型殘差分析之結果 127 附錄E 各模型自變數之VIF值 14