應用類神經網路預測發電廠周圍細懸浮微粒及臭氧濃度

Predicting Concentrations of Fine Particulate Matter and Ozone Around the Poewer Plants Using Artificial Neural Network

張家維(Jia-Wei Jhang)

指導教授 ： 王玉純

中原大學/工學院/環境工程學系/碩士(2021年)

https://doi.org/10.6840/cycu202100448

摘要

根據2020年統計燃煤占發電量的36%，表示臺灣燃煤發電量仍為高占比，而典型預測空氣品質模式包含高斯類擴散模式、軌跡類模式與網格類模式等，本研究以機器學習中的類神經網路(Artificial Neural Network, ANN)整合分析環保署固定污染源連續自動監測設施資料(Continuous Emission Monitoring Systems, CEMS)、環保署空氣品質監測資料、中央氣象局氣象觀測資料，用以預測鄰近三座台電電廠(林口、台中、大林)每日的細懸浮微粒與臭氧濃度並探討預測結果。
蒐集2019至2020年林口、台中、大林電廠之固定污染源每小時排放數據(包含氮氧化物及二氧化硫)，林口、沙鹿、小港空氣品質監測站之每小時空氣污染物數據(包含二氧化硫、二氧化氮、一氧化碳、臭氧、細懸浮微粒及懸浮微粒)與林口、梧棲、鳳森氣象觀測站之每日大氣數據(包含溫度、氣壓、濕度、風速及風向)等參數，將前述資料彙整為日平均值導入以Python語言程式建構的ANN模型中作為輸入參數，輸出參數為細懸浮微粒(μg/m3)與臭氧(ppb)，決定係數(Coefficient of determination, R2)、均方根誤差(Mean Square Error, MSE)與絕對平均誤差(Mean Absolute Error, MAE)用以評估模型表現，隱藏層神經元試誤後，結果得出細懸浮微粒與臭氧分別使用22個與23個隱藏層神經元有最佳的預測表現，再將數據分區建立模型，保留訓練與驗證時連接層與層之間的最佳權重值，用以預測2021年1月至3月的細懸浮微粒與臭氧濃度。
以ANN預測每日細懸浮微粒濃度，林口空氣品質監測站之最佳設定在丟棄神經元為 20%，學習速率 0.01平均R2值為 0.9057；沙鹿空氣品質監測站之最佳設定在丟棄神經元為 0%，學習速率 0.004平均R2值為 0.9243；小港空氣品質監測站之最佳設定在丟棄神經元為 0%，學習速率 0.004平均R2值為 0.9303。以ANN預測每日臭氧濃度，林口空氣品質監測站之最佳設定在丟棄神經元為 20%，學習速率 0.006平均R2值為 0.7523；沙鹿空氣品質監測站之最佳設定在丟棄神經元為 0%，學習速率 0.008平均R2值為 0.7437；小港空氣品質監測站之最佳設定在丟棄神經元為 20%，學習速率 0.01平均R2值為 0.7626。預測結果顯示，細懸浮微粒預測值低於實際值，結果與訓練模型時結果大致相符；然而臭氧預測值大部分高於實際值，模型預測細懸浮微粒表現較佳。
本研究建立預測細懸浮微粒之ANN模型，僅使用少部分參數及資料量即可達到良好的預測效果，可以作為未來預測空氣品質的參考依據，建議未來能夠應用於其他電廠進行預測評估。此外，臭氧結果較差原因推測臭氧多為衍生性，並非由污染源直接排放，建議未來臭氧預測納入光化學評估監測站、揮發性有機物等相關臭氧前驅物之數據與傳輸、沉降、排放等大氣因子作為建立模型之參數。

關鍵字

火力發電廠 ； 類神經網路 ； 細懸浮微粒 ； 臭氧

並列摘要

According to 2020 statistics, 36% of electricity generation was from coal-fired industries indicating a higher proportion of coal-fired power generation in Taiwan. Typical predictive models for air quality monitoring include Industrial Source Complex Short-Term Dispersion Model (ISCST3), Trajectory model, and Grid model etc. This study used Artificial Neural Network (ANN) in machine learning integrating analysis of Continuous Emission Monitoring Systems (CEMS), air quality monitoring data, and meteorological data to predict the ambient concentrations of fine particulate matter (PM2.5) and O3 nearby 3 power plants, namely Linkou, Taichung, and Dahlin.
The study collected hourly data of NOx and SO2 from CEMS in power plants, the air pollutant data (including SO2, NO2, CO, O3 PM2.5, PM10) from Linkou, Shalu and Xiaogang air quality monitoring stations, and atmospheric data (including temperature, pressure, humidity, wind speed, wind direction) from Linkou, Wuqi, and Fongsen meteorological observation stations from 2019 to 2020. The hourly data was aggregated as daily averages and imported in to the ANN model built in Python environment as input parameters. Output parameters are daily concentrations of PM2.5 (μg/m3) and O3 (ppb), and R2 (Coefficient of determination), MSE (Mean Square Error) and MAE (Mean Absolute Error) were used to assess the model performance. After the hidden layer neuron test, the best predictions were obtained as 22 hidden layer neuron for PM2.5 and 23 for O3. Finally, the study predicted the daily concentrations of PM2.5 and O3 using the best weights among input-hidden-output layers for the training period from January to March 2021.
The results of the prediction of PM2.5 from training models identified the best setting with a dropout of 20%, learning rate of 0.01, and with average R2 of 0.9057 in Linkou air quality monitoring station; a dropout of 0%, learning rate of 0.004, and with average R2 of 0.9243 in Shalu air quality monitoring station; and a dropout of 0%, learning rate of 0.004, and with average R2 of 0.9303 in Xiaogang air quality monitoring station. Predicted O3 results also showed the best setting with a dropout of 20%, learning rate of 0.006, and with average R2 of 0.7523 in Linkou air quality monitoring station; a dropout of 0%, a learning rate of 0.008, and with average R2 of 0.7437 in Shalu air quality monitoring station; and a dropout of 20%, learning rate of 0.01, with average R2 of 0.7626 in Xiaogang air quality monitoring station. In the forecast results for January to March 2021, predicted values are lower than actual values for the PM2.5 that are broadly in line with those obtained during the training model, however, most of the predicted O3 values are higher than the actual values. ANN models had better predict performance for ambient PM2.5.
The study applied the ANN model for predicting fine particulates that utilized only a small number of parameters and data quantity to achieve good prediction results. Therefore, it can be used as a reference for future air quality prediction. Furthermore, this study suggested that the ANN can be applied to other power plants for future projections and evaluation. In addition, the reason for the poor O3 predictions can be presumed because O3 is mostly derivative and not directly emitted by the source. Hence, the study proposes future ozone predictions should include data from Photochemical Assessment Monitoring Stations, ambient concentrations of volatile organic compounds, and other atmospheric factors, such as transport, deposition, and emissions as parameters for modeling.

並列關鍵字

Power Plants ； Artificial Neural Network ； Particulate matter ； Ozone

參考文獻

1. Zhihua (Tina) Fan, Q.M., Clifford Weisel, Robert Laumbach, Pamela Ohman-Strickland, Stuart Shalat, Marta Z Hernandez Kathleen Black Acute exposure to elevated PM2.5 generated by traffic and cardiopulmonary health effects in healthy older adults. 2008.

2. 王建楠 and 李璧伊, 細懸浮微粒暴露與心血管疾病: 系統性回顧及整合分析. 中華職業醫學雜誌, 2014. 21(4): p. 193-204.

3. Lim, C.-H., et al., Understanding global PM2.5 concentrations and their drivers in recent decades (1998–2016). Environment International, 2020. 144: p. 106011.

4. 行政院環境保護署-國家溫室氣體減量法規資訊網, 能源新配比 經長保證不缺電. 2019.

5. 台灣電力公司.