評估整合多重時間解析度受體模式及限縮受體模式之成效: 數值模擬研究

Abstract

精準的辨認並且評量細懸浮微粒(Fine particulate matter和揮發性有機物(volatile organic compounds)這些危害物的汙染源，是非常重要的。化學平衡法(Chemical mass balance)、絕對主成分分析(absolute principal component scores)和正矩陣因素分析(positive matrix factorization)是已知，並且被提出可以解決上述問題的方法。這三種方法，被稱作受體模式。 化學平衡法、絕對主成分分析和正矩陣因素分析有各自的使用時機。然而，在資料組的時間不統一，而且掌握的先驗資訊不完整的情況下，並沒有一個合適的受體模型，可以精準的分析汙染源。因此，本研究的目的，就是驗證結合多重時間解析度正矩陣因素分析以及限縮正矩陣因素分析的模式，分析汙染源的精準程度。 本研究經由數值模擬、模式執行、模式成效評估與實地資料應用為主。數值模擬中的以自然對數常態分佈生成小時值，以美國環保署的資料庫(US EPA SPECIATE dataset)為生產來源。模型執行使用多線性引擎(Multilinear Engine)軟體來執行新舊兩個模型。資料分析以平均絕對值誤差為成效評估標準(average absolute error)。實地資料則採自台北萬華測站。 單純多重時間解析度受體模式與整合模式的AAEf依序為0.18與0.13；AAEg依序為0.11與0.10；決定係數依序為0.68與0.78。實地採樣的資料若使用單純多重時間解析度受體模式，可以被解析出六個汙染源，分別是汽車排放一、區域傳播、工業塗料、汽車排放二、天然氣以及二次硫化物。實地採樣的資料若使用整合受體模式，則可以被解析出七個汙染源，分別是汽車排放一、區域傳播、工業塗料、汽車排放二、天然氣、二次硫化物以及揚塵。這個結果與數值模擬的結果相符合。整體而言，不論是汙染源隨時間貢獻量的預測，還是汙染源化學組成的預測成效，整合模式皆優於單純多重時間受體模式。

Fine particulate matter (PM2.5) and volatile organic compounds (VOCs) have been well known to relate to adverse health effects. Thus, to precisely identify and to evaluate the sources of both PM2.5 and VOCs are important. Chemical mass balance (CMB), absolute principal component scores (APCS), and positive matrix factorization (PMF), have been used to attend the above purposes. These three models can be defined as receptor models. CMB, APCS and PMF have been used in different situations. However, they usually do not work for the situation that the data has multiple time resolution and the factor profile or the source contribution is incomplete. Thus, an evaluation of the performance of a combination of the constrained PMF model and a multiple-time-resolution PMF model was conducted in this study. A synthetic data was used in this study to achieve the above purposes. This study was conducted through the method of data simulation, model implementation, evaluation of the performance of models and application to field study. The synthetic data was created as an hourly measurement contribution matrix, where the hourly fluctuation was added and was assumed to be random and log-normal distribution. Six source profiles, which included petroleum refinery, vehicle exhaust, industrial coating, coal combustion, natural gas and fugitive dust were created from US EPA SPECIATE dataset. Two models were run in this section so that two models could compare with each other. One was a multiple-time-resolution model, and the other was a mixed model which combined a multiple-time-resolution model and a constrained model. These two models were compared to each other by the average absolute error. The data of field study was collected from the Wanhua monitoring site in Taipei. The AEEf of the multiple-time-resolution-only receptor model and of the combined receptor model are 0.18 and 0.13, respectively; The AAEg of the multiple-time-resolution-only receptor model and of the combined receptor model are 0.11 and 0.10, respectively; The R square are 0.68 and 0.78, respectively. For a conclusion, the performances of the combined model in the prediction of the factor profile and in the source contribution were better than the performance of the multiple-time-resolution-only model.