建立大數據深度學習模式以推估區域性多時刻空氣品質-以高雄市為案例

Building a deep learning model with big-data to estimate regional multi-time air quality - a case study of Kaohsiung City

吳冠諺(Kuan-Yen Wu)

指導教授 ： 張斐章

共同指導教授 ： 潘述元(Shu-Yuan Pan)

國立臺灣大學/生物資源暨農學院/生物環境系統工程學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202102471

摘要

空氣污染為當今人們所關注的議題之一，空氣污染物包含了PM2.5、PM10、O3、SOX、NOX等，其中，以PM2.5對人體造成的危害性最大，環保署於前幾年陸續在工業區附近廣泛設置微型感測器，而臺灣PM2.5預測模式大多為預測環保署測站PM2.5濃度，並藉由空間內插等方式，以輸出區域性的預測結果，然而工業區附近空污的濃度通常較高並且變化大，使得大多數預測模式對於此區域的預測結果仍然有進步之空間，故本研究建立一推估模式(AE-CNN-BP)，由數種ANN模式所組成，包含了自編碼器(AE)、卷積神經網路(CNN)、倒傳遞神經網路(BPNN)，並用於推估T+X時刻高雄市的PM2.5濃度。推估模式基於前人時間序列預測模式於環保署測站預測T+X時刻PM2.5濃度的結果，作為CNN-BP模式的輸入值，以預測編碼(Code)，並將此預測結果作為解碼器(Decoder)的輸入值，即可推估環保署與微型感測器測站T+X時刻之PM2.5濃度值。本研究結果顯示，AE模式在訓練及測試階段之準確度都相當高，RMSE介於8~9.2(μg/m3)之間，R2介於0.93~0.94之間，在預測編碼上，總共建置了三種預測模式，分別為倒傳遞神經網路(BPNN)、深層神經網路(DNN)、混合卷積神經網路(CNN-BP)，結果顯示，CNN-BP模式在預測編碼上獲得最高的準確度；而AE-CNN-BP推估模式應用於高雄市的工業區，結果顯示，在高污染時段下，其PM2.5濃度推估結果較前人研究更為準確，並且在中、高污染時段，更能發揮空污預警之效用，故本研究證明了微型感測器應用於空污預測和預警上是具有價值的。

關鍵字

PM2.5預測 ； 深度學習 ； 微型感測器 ； 自編碼器 ； 卷積神經網路 ； 倒傳遞神經網路 ； 臺灣

並列摘要

Nowadays, air pollution is one of the popular topics of concerns. Air pollutants consist of PM2.5, PM10, O3, SOX, NOX, and others. Among them, PM2.5 is the substance the most hazardous to human beings. In the past few years, the Environmental Protection Administration in Taiwan (TW EPA) has set up a wide range of air quality micro-sensors near industrial areas. However, most of the PM2.5 forecast models in Taiwan focus on the point forecasts of PM2.5 concentration at the air quality monitoring stations established by TW EPA. The regional forecasts of PM2.5 concentration can be obtained by applying spatial interpolation methods on the aforementioned point forecasts. Nevertheless, air pollution at industrial areas is usually severe with high variation in concentration. Therefore, there is still room for air quality forecast models to improve accuracy at industrial areas. This study builds an estimation model (AE-CNN-BP) that comprises several artificial neural network (ANN) models including Autoencoder (AE), Convolutional Neural Network (CNN), and Back Propagation Neural Network (BPNN) for estimating multi-step-ahead PM2.5 concentration in Kaohsiung City of Taiwan. The CNN-BP of the proposed model is to predict the code for use in AE based on the multi-step-ahead PM2.5 forecast results of our previous study at air quality monitoring stations. Then, the predicted code will be decoded to estimate the multi-step-ahead PM2.5 concentrations of air quality micro-sensors. The results show that the AE model achieved high accuracy in training and testing stages, with RMSE falling within 8 μg/m3 to 9.2 μg/m3 and R2 falling between 0.93 and 0.94. In order to predict the code, three different ANN models were built for comparison purpose, which were BPNN, Deep Neural Network (DNN), and CNN-BP models. The predictions of the code show that the CNN-BP model had the highest accuracy. These results indicated the proposed AE-CNN-BP estimation model not only performed more accurately than the previous study in the periods of high pollution but also could benefit the issuance of air pollution warnings in the periods of moderate and high pollution. In sum, this study demonstrates the applicability of air quality micro-sensors in air pollution forecasts and warnings.

並列關鍵字

PM2.5 forecast ； Deep learning ； Micro-sensor ； Autoencoder ； Convolutional Neural Network (CNN) ； Back Propagation Neural Network (BPNN) ； Taiwan

參考文獻

1. ApSimon, H., Oxley, T., Woodward, H., Mehlig, D., Dore, A., Holland, M. (2021). The UK Integrated Assessment Model for source apportionment and air pollution policy applications to PM2.5. Environment International, 153, 106515.

2. Apparicio, P., Gelb, J., Carrier, M., Mathieu, M. È., Kingham, S. (2018). Exposure to noise and air pollution by mode of transportation during rush hours in Montreal. Journal of Transport Geography, 70, 182-192.

3. Abirami, S., Chitra, P. (2021). Regional air quality forecasting using spatiotemporal deep learning. Journal of Cleaner Production, 283, 125341.

4. Brauer, M., Casadei, B., Harrington, R. A., Kovacs, R., Sliwa, K., WHF Air Pollution Expert Group. (2021). Taking a Stand Against Air Pollution—The Impact on Cardiovascular Disease: A Joint Opinion from the World Heart Federation, American College of Cardiology, American Heart Association, and the European Society of Cardiology. Circulation, 143(14), e800-e804.

5. Bai, Y., Li, Y., Wang, X., Xie, J., Li, C. (2016). Air pollutants concentrations forecasting using back propagation neural network based on wavelet decomposition with meteorological conditions. Atmospheric pollution research, 7(3), 557-566.