113

3,950 千元

數位建設-建構開放政府及智慧城鄉服務

2024/04/01

2024/12/31

蔡春進

環境部監測資訊司

黃健瑋

台灣PM2.5監測與控制產業發展協會

大氣空氣、環境監測及檢測

PM2.5感測器、低成本空氣感測器、監測、環境物聯網、多變數迴歸

PM2.5 sensor, low cost air sensor, monitoring, Environmental internet of things, Multiple variable regression

洪廣華

(03)5712121#55549

ghhong.ev05g@nctu.edu.tw

本計畫工作內容為巡檢已布建在18縣市的PM2.5感測器性能，以及針對VOC感測器、異臭味感測器測值的不穩定性提出改善方案，希望提升我國感測器之數據品質，對於感測器物聯網的推動及使用有所助益。本年度截至民國113年11月已完成607台PM2.5感測器的查核，占全體設置數量的5.8%。查核範圍包含台灣本島17個縣市的現地查核，以及外島連江縣的線上遠程查核，各縣市的感測器數據品質滿意度均超過90%，顯示各縣市布建的感測器具有相當好的數據品質，數據準確性與穩定性方面表現優良。 在VOC感測器測值準確性提升之研究項目，本團隊已成功運用VOC感測器之影響性因子及相關性因子開發校正模型，透過線性迴歸(Linear Regressio, LR)及機器學習(Machine Learning, ML)等校正方法提升VOC感測器的數據品質。於實場試驗，XGBoost模型有著最好的校正成效，添加T、RH及CO等影響性因子為校正參數，Sensirion SGP 30感測器可由原廠測值的決定係數R平方 (R-squared，R2) (0.44)、正規化誤差 (Mean Normalized Error，MNE) (156%) 及均方根誤差 (Root Mean Square Error，RMSE) (0.091 ppm) 提升至R2 (0.93)、MNE (23%)、RMSE (0.017 ppm)，而再添加相關性因子NOx作為校正參數後，能進一步提升數據品質，提升至R2 (0.96)、MNE (16%)、RMSE (0.013 ppm)，符合美國環保署規範的補充監測網等級 (MNE<20%)。另外，本計畫發現SHARP ms-1100 感測器與O3存在交叉敏感性，利用XGBoost模型將其作為校正參數能夠使感測器原廠測值的R2 (0.38)、MNE (294%)、RMSE (0.166 ppm) 提升至R2 (0.96)、MNE (22%)、RMSE (0.016 ppm)，符合美國環保署規範的熱點識別與污染鑑識等級 (MNE<30%)。本計畫利用不同校正方法及校正參數開發校正模型，成功提升另一年份感測器的測值準確性，也成功將模型運用於污染熱區的感測器測值校正，土城測站方圓一公里內的六台感測器R2皆提升至0.70以上，其中有五台感測器的MNE符合美國環保署規範的熱點識別與污染鑑識等級 (MNE<30%)，證實該模型具有長效性及實用性。本計畫開發之校正模型可改善臺灣VOC感測器測值高估的問題，以幫助縣市環保局追蹤VOC的排放源。 在異臭味氣體感測器影響性因子之研究項目方面，本計畫發現感測器在實場的測值受溫濕度影響很大，NH3與H2S的測值高估很多。本研究使用加熱法，在實驗室測試時可使Figaro EC NH3(FECS44-100)和H2S(FECS50-100)感測器的腔體內相對溼度降低至穩定的40%至48%範圍之間。在NH3方面，FECS44-100 NH3感測器的實驗室測值(15分鐘平均值)和參考儀器作線性比對，加熱前原斜率為0.72、截距為280.9ppb、R2為0.72、RMSE為397.23ppb及 NRMSE為39.92%，加熱及大氣背景值校正後分別改善為斜率0.86、截距119.44ppb、R2 0.86、RMSE 204.53ppb及 NRMSE 23.94%；在H2S方面，FECS50-100 H2S感測器的實驗室測值(1分鐘平均值)和參考儀器作線性比對，加熱前原斜率為0.39、截距為75.17ppb、R2為0.39、RMSE為40.25ppb及NRMSE為32.42%，加熱及大氣背景值校正後分別改善為斜率0.95、截距5.71ppb、R2 0.95、RMSE 14.83ppb及NRMSE 13.7%。此實驗室結果顯示兩種異味感測器的數據品質皆有所提升。在實場測試方面，養豬場的現場NH3濃度範圍為445.6±321.3ppb時，FECS44-100 NH3感測器測值(15分鐘平均值)和參考儀器作線性比對，加熱及大氣背景校正後的R2由原本的0.03改善為0.15；但當現場NH3濃度較低時(<100ppb)校正較為困難，仍有持續改善的必要。本研究進一步對養雞場的DD Scientific NH3感測器進行NLR校正並以濕度作為參數，將NH3與PPWD-IC的測值(15分鐘平均值)進行比對，NLR校正後的R2由原本的0.13改善為0.42， RMSE由原本的543.0ppb大幅降低至54.4ppb，顯示NLR校正公式在可以提高DD Scientific NH3感測器的數據品質準確性，且濕度為影響NH3的重要參數。另外，FECS50-100 H2S感測器在各個實場的校正結果皆改善許多，H2S感測器測值(1分鐘平均值)和參考儀器作線性比對，加熱及大氣背景校正後的R2可以由原本的0.03至0.04改善為0.43至0.59，數據品質大幅提升。

PM2.5感測器、低成本空氣感測器、監測、環境物聯網、多變數迴歸

MOENV054113012.pdf

23MB

2025/1/24 上午 12:00:00

The work content of this project involves inspecting the performance of PM2.5 sensors deployed in 18 counties and cities, as well as proposing improvement plans to address the instability of VOC sensors and odor sensors, with the aim of enhancing the data quality of sensors in Taiwan and contributing to the promotion and use of sensor-based IoT systems. As of November 2024, a total of 607 PM2.5 sensors have been inspected, accounting for 5.8% of all installed sensors. The data quality satisfaction rate for sensors in all counties and cities exceeded 90%, demonstrating that the deployed sensors exhibit excellent data quality, accuracy, and stability. Research project on improving the accuracy of VOC sensor measurements: Our team has successfully developed calibration models using the influencing factors and related factors of VOC sensors. By applying calibration methods such as Linear Regression (LR) and Machine Learning (ML), we have improved the data quality of VOC sensors. In the field test, XGBoost models shows the best calibration performance. By adding influencing factors such as T, RH, and CO as calibration parameters, the Sensirion SGP 30 sensor's manufacture calibrated measurement coefficients of determination (R2), Mean Normalized Error (MNE), and Root Mean Square Error (RMSE) improve from 0.44, 156%, and 0.091 ppm to 0.93, 23%, and 0.017 ppm, respectively. Further improvements are achieved by adding the correlation factor NOx as a calibration parameter, reaching R2 (0.96), MNE (16%), and RMSE (0.013 ppm), complying with the US EPA's Supplementary Monitoring Network level (MNE<20%). Additionally, this study discovers cross-sensitivity between the SHARP ms-1100 sensor and O3. When the O3 was used as a calibration parameter, the XGBoost model can improve the sensor's manufactured calibrated measured values of R2, MNE, and RMSE from 0.38, 294%, and 0.166 ppm to 0.96, 22%, and 0.016 ppm, respectively, complying with the US EPA's Hotspot identification and Pollution identification levels (MNE<30%). This study successfully improves the measurement accuracy of sensors from another year using different calibration methods and parameters and applies the model to calibrate sensor measurements in pollution hotspots. Within a one-kilometer radius of the Tucheng AQMS, the R2 of six sensors improved to above 0.70, with five sensors complying with the US EPA's Hotspot identification and Pollution identification levels (MNE<30%), confirming the model's long-term effectiveness and practicality. The calibration model developed in this study can solve the problem of overestimation of VOC sensor values in Taiwan and assist county and city EPB in tracking VOC emission sources. This study uses heating and humidity control methods to maintain the relative humidity in the Figaro EC NH3 and H2S sensor chambers between 40% and 48%, reducing changes in temperature and humidity. FECS44-100 NH3 sensor heating background value can be reduced to -0.23 to 0.19 mV through ΔAH correction, the measured value jump range is 308.6 ppb, LOD is 179.3 ppb. Unheated background value can be reduced to -0.65 to 0.5 mV through MLR-Ta and RHa correction. The measured value jump range is 554.11 ppb, and the LOD is 441.67 ppb. Heating measurement value compared with the unheated measurement value , the slope increased from 0.72 to 0.86, the intercept decreased from 280.9 ppb to 119.44 ppb, R2 increased from 0.72 to 0.86, RMSE and NRMSE were reduced to 204.53 ppb and 23.94%, and the data quality was improved. The background value of the FECS50-100 H2S sensor is not affected by the ambient temperature and humidity. Compared with the heated measurement value and the unheated measurement value, R2 increased from 0.39 to 0.95, and the intercept dropped from 75.17 ppb to 5.71 ppb, RMSE dropped from 40.25 ppb to 14.83 ppb and NRMSE dropped from 32.42% to 13.7%. In field tests conducted at a pig farm, the ambient NH₃ concentration ranged from 445.6 ± 321.3 ppb. The FECS44-100 NH₃ sensor readings (15-minute averages) were linearly compared with those of the reference instrument. After heating and atmospheric background corrections, the R² improved from 0.03 to 0.15. However, when NH₃ concentrations were relatively low (<100 ppb), calibration remained challenging and required further improvement. This study further applied NLR (Non-Linear Regression) method to calibrate the DD Scientific NH₃ sensor at chicken farm, using humidity as a parameter. The NH₃ sensor readings (15-minute averages) were compared with those from the PPWD-IC reference instrument. After NLR calibration, the R² was improved significantly from 0.13 to 0.42, and the RMSE was decreased substantially from 543.0 ppb to 54.4 ppb. These results demonstrate that the NLR calibration model effectively enhances the data accuracy of the DD Scientific NH₃ sensor and highlights humidity as a critical influencing factor for NH₃ measurements.

PM2.5 sensor, low cost air sensor, monitoring, Environmental internet of things, Multiple variable regression