都會地區PM2.5監測系統與預測之研究

Abstract

近年來，隨著環境與健康意識抬頭，都會地區空氣汙染的議題逐漸備受矚目，尤其是造成人體健康危害最嚴重的PM2.5¬。許多研究皆已證實空氣汙染物存在空間變異性，間接證明居民之暴險濃度值與周界環境濃度值之間具有顯著差異，遑論周界環境濃度值的代表性。有別於佈署大規模之空氣盒子靜態測站，或是攜帶儀器之移動式定點測站，用以精準量測居民之暴險濃度，本研究開發出一以物聯網技術為基礎的車載監測系統(On-board monitoring system, OBMS)，負責於都會地區進行PM2.5之即時監測，量測數據包含當前時間、位置、移動速度、PM1、PM2.5、PM10、環境溫度與濕度，不僅能達到靜態測站的精準測量，亦能經濟且效率地測量空間中的濃度變化，尤其是交通汙染與社區汙染源在都會地區中對於居民暴險濃度的影響。 除了即時監測暴險濃度之外，都會地區空氣品質預測之重要性亦被廣泛認可，準確的空氣品質預測使得政府與民眾足以提前採取相關預防或是減緩的措施。受限於準確的汙染源資料取得不易、汙染物交錯複雜的化學反應、簡化的汙染物傳輸模式難以實際量化真實情況，許多傳統的預測模型難以突破上限，因此，擅長從大數據當中找尋模式與規則的機器學習方法近年來蓬勃發展，唯使用機器學習方法的限制在於資料本身的多寡優劣，許多使用機器學習作為預測方法的文獻當中，不僅資料來源採用代表性不足的周界環境濃度值，時間預測刻度多半也以小時計，對於居民暴險濃度來說，代表性、即時預測性實為欠缺。因此，本研究以機器學習方法中的遞迴神經網路-長短期記憶網路(Long short-term memory, LSTM)與其重要變形版本(Gated recurrent unit, GRU)作為預測方法，搭配空氣盒子即時數據作為主要的資料來源，以台北都會區作為研究範圍，發展出一時間刻度更細緻的預測模型，包含預測未來6分鐘至30分鐘的濃度變化，並結合OBMS蒐集而得的數據進行預測比對。本研究結果證實，藉由車載監測系統即時監測空間尺度更細緻的居民暴險濃度與時間尺度更細緻的空氣品質預報乃是可行的，藉由物聯網與機器學習的方式，實現大數據蒐集與分析預測，提供決策者與民眾更精準的依據來源，漸漸邁向智慧城市的目標。

Recently, with the rise of environmental and health consciousness, issues about air pollution in urban areas are gradually attracting attention, especially for PM2.5 that causes the most severe health risk to human beings. Many studies have highlighted the phenomenon of spatial variability in air pollution, indirectly supporting that there are significant differences between residents’ exposure and background levels. Based on such a factor, background levels cannot represent residents’ exposure around a static station. Different from deploying large-scale airbox stations or mobile fixed-point stations with portable instruments to accurately measure residents’ exposures, this study therefore develops an on-board monitoring system (OBMS) based on the IoT technology, which is responsible for real-time PM2.5¬ measurement in urban areas. In addition to PM¬2.5¬, the measured parameters include current time, current location, moving speed, PM1, PM10, ambient temperature and humidity. The OBMS not only achieves the goal of accurate measurement like what static stations can do, but also efficiently measures the PM2.5 levels in space in an economical way, especially that its measuring results can indicate the impacts of traffic pollution and community pollution in urban areas on residents’ exposure. In addition to providing real-time monitoring data on residents’ exposures, it is also important to predict air quality in urban areas based on these kind of monitoring data (however, limited by small data collected by OBMS, the prediction is based on large data collected by airbox stations). Accurate air quality prediction enables government officials and the public to take proper prevention actions in advance. However, it is difficult to accurately evaluate a real situation because accurate pollution data are hard to obtain, chemical reactions of pollutants are complicated, and the ways utilized to describe pollutant transmission are too simple. Thus, many traditional prediction models have their own limitations, making many prediction tasks use statistical methods that are good at finding patterns and rules from big data. Machine learning is one of statistical methods, which has been adopted by many air quality prediction studies. In these studies, data sources generally use under-represented background levels, and the prediction intervals are mostly in hours, which may not be suitable for air quality prediction needed by residents. Therefore, this study proposes to use long short-term memory (LSTM) and gated recurrent unit (GRU) as the prediction method. With the real-time airbox data as the main source of data, the Taipei metropolitan area is used as a research area to develop a more detailed forecasting model in which the PM2.5 levels changes in next 6 to 30 minutes are predicted. The predictions are also compared with the data collected by the OBMS. This study confirms that the OBMS can instantly monitor residents’ PM2.5 exposure with smaller time interval, and that using smaller time intervals to predict air quality is feasible. Using the techniques of Internet of Things and machine learning, big data collection, analysis and forecasting can be achieved, decision-makers and people can obtain more accurate information, and the goal of smart cities can be met eventually.