EPA-106-L102-02-A142

106

2,950 千元

環境監測資訊-空氣品質監測規劃與測站管理

2017/02/13

2017/12/31

余化龍

監資處

林書庸

國立臺灣大學

環境監測及檢測

空氣感測器，可信度分析，時空特徵分析

Air quality sensor, credibility analysis, space-time features analysis

胡明哲,林遠見

陳以容

02-33663454#

r89622042@ntu.edu.tw

因應空氣污染問題日益受到重視，行政院環保署計畫於臺灣架設萬點空品感測器，搭配環境物聯網以改善未來環境空氣品質預測與管理之應變能力。目前，臺灣市場上感測器的種類包含了Edimax空氣盒子、LASS感測器、華碩簡易型空品感測器等。隨著感測器佈點增加及環境物聯網的發展，民眾可透過感測器得知時空分布更細微的PM2.5濃度。本團隊使用感測器所量測的PM2.5濃度觀測數據作為數據分析及調校模型的建立基準，但因感測器本身設計測量原理及易受環境溫濕度及風力影響，在量測PM2.5濃度上具有限制性及較大的量測誤差，此量測誤差易造成民眾恐慌或質疑環保署監測站數據的真實性。因此，本計畫為避免因感測器數據產生的誤差值造成民眾的恐慌，利用系統性流程針對感測器數據進行品質檢核與調校，包括感測器異常值分析及調校模型建立，以提供符合真實值的感測器數據。除此之外，本團隊除了在Edimax空氣盒子上進行異常值剔除，亦針對了LASS感測器的數據進行異常值剔除。 本團隊由雲端資料取得感測器數據進行資料檢核(由工研院取得，數據期間105/10/15-106/3/15)及調校模型建立(由環保署環境物聯網平臺取得，數據期間106/6~106/12)；在資料檢核方面，利用時間及空間離群值檢測方法，可偵測空污突發事件及剔除感測器離群值，提供更符合實際的資料與資訊；在調校模型建立方面，主要考量臺灣各地區之環境條件不一，分別針對北、中、南三個地區進行並探討感測器與環保署監測站PM2.5差值之間的關係，其結果發現，在感測器測值介於20~40 μg/m^3時，感測器與標準測站之間的差異非常小，約在10 μg/m^3以內，若感測器測得濃度高於40 μg/m^3及低於20 μg/m^3，則感測器與標準測站間的差值會顯著的提高。 由於感測器及環保署監測站數據不管在空間或者時間的解析度及特性都不一樣，資料融合推估技術上考量了以上的問題，並利用校正後感測器數據與其附近環保署監測站資料進行推估及融合(數據校正時間範圍106/6~106/12)，以提供更完整的空品資訊。再利用融合後之空品資料，進行時空特徵萃取技術應用，找出各地區的PM2.5的時間及空間的分佈。 本團隊在時空特徵萃取技術上以臺中地區為例，將融合後的PM2.5數據萃取出日與周的時間及空間特徵。在沿海地區發現其日趨勢顯示具有上午及下午兩個波鋒，有較高的PM2.5濃度變化；在周週期的時空間特徵變動趨勢，發現在週五時刻有較高的PM2.5濃度變化，周末有逐漸下降趨勢。在萃取特徵中，透過時空特徵萃取技術，能找出不同空間及不同時間下之PM2.5濃度變化，並進一步提供空氣污染管控及防治之資訊。 本團隊針對精緻空品預報規畫以北部地區為主要區域，建立PM2.5濃度與空氣污染排放、土地利用、路網長度及其他氣象資訊之間的關係，共2233個變數，並參酌了美國及中國的精緻空品預報技術，可作為未來精緻空品預報規畫之參考。 另將空氣品質監測值時空資訊進行視覺化平臺開發，並研發即時動態儀表板，讓空品時空巨量資料得以快速及簡易地讓使用者及決策者即時了解空品現況。 由於感測器比對作業因時程關係僅比對部份時間，且參數部份還考慮PM2.5形成機制。故建議延續蒐集數據量測時間以進行較完整的評估，本融合資料推估技術及成果，可結合健康風險分析，後續可發展相關健康風險評估地圖。

空氣感測器，可信度分析，時空特徵分析

環保署感測器計畫期末報告-version12.pdf

12MB

2018/12/31 上午 12:00:00

With the increasing attention paid on air pollution issues in Taiwan, the Environment Protect Admiration has plan to set up more than ten thousand sensors in Taiwan to improve the resilience of future air quality forecasts and management. There are many types of sensor on Taiwan market including Edimax, LASS, Asus and etc. With the increase of sensors and the development of internet, people can assess information about PM2.5instantly. However, due to the design and diverse environment condition, there are limitation on PM2.5observation. Therefore, the purpose on this project are develop mathematical calibration model to conduct the quality inspection on sensor data, including analysis of the outlier of sensor data and its calibration model. Due to the vary on environment condition in Taiwan, we have developed calibration model on three regions including north, central and southern part of Taiwan to discuss the relationship between the difference of sensor and EPA station monitor data on PM2.5. When the sensor data observe 20-40μg/m^3, the difference is very small; while when the sensor data is below or above the difference, the difference will be significantly increased. In the quality inspection on sensor data, we develop 3 methods including temporal outlier, spatial outlier and spatio-temporal buffer outlier detection to consider the different of space time data of sensor in Taiwan. In temporal outlier, we able to detect the error cause by machine that observe abnormal data and sudden air pollutant events. Spatial outlier detection determines the sensor in a different environment; spatiotemporal buffer outlier detection method provides a comprehensive type of outlier Value detection and proposed. Through the above three methods and sensor data, the program successfully detects and proposes sensor outliers. Besides that, we also conduct data fusion estimation techniques on EPA monitor data and sensor data. Due to the different resolutions and characteristics of sensors and EPA stations in space or time, the data fusion estimation technique considers the above problems and uses the corrected sensor data and its nearby EPA monitoring stations data to estimate and integration, to provide more complete information on air pollutant. We also applied spatio-temporal feature extraction technique on air pollutant data to find the distribution of PM2.5.in Taiwan. In the spatio-temporal feature extraction technique, the PM2.5 data were extracted from the time and space characteristics of day and week to understand the changes of air quality in different regions. We identify the activities such as work commuting time in various regions in daily cycle feature and parks and camps affect the PM2.5. change in the region due to human activities due to weekend time on weekly cycle. In addition, the program also succeeded in identifying air pollution incidents in the extraction characteristics. Through spatio-temporal feature extraction techniques, PM2.5concentrations in different spatial and time periods can be identified and the information on air pollution control and prevention can be provided. In 2011, we had finish the retrospective of spatio-temporal PM2.5 distribution Taipei. Since there was no sensor observation data at that time, only few EPA monitoring stations data were used to develop model with land use data. In this research plan, we use north Taiwan as example. Beside explore the relationship between PM2.5 and land use data, we also include road length, air pollution emissions and climate data to establishes the relationship between various environmental variables and PM2.5. Through above data, we had developed 2233 of different variables, and explore the relationship between 2233 variables and PM2.5. We also develop visualization of space-time information on air quality monitoring and developed dynamic dashboard to enable users and decision makers to instantly understand the status of air and space products quickly and easily. Due to the work alignment, the work on the sensor data comparison operation, we still need to consider the formation of PM2.5 in our quality inspection process. For a complete assessment, we hope to continue the project. Consequential results of the fusion data can be followed up with the results of the health risk analysis to develop a map of the relevant health risk assessment.

Air quality sensor, credibility analysis, space-time features analysis