| 摘要: | 研究早已表明空氣污染對於人體健康有著嚴重的影響，然而由於空氣污染物存在高度非線性與非穩態的特徵，而且影響空氣污染物的各種因素與化學反應非常複雜，現今仍沒有一套完整的空氣品質分析流程與空氣品質預測模式。因此，本研究以資料探勘的角度，從實際觀測資料分析出發，結合交通因子，大氣因子與水文條件進行分析與降雨特性水文分析並提出完善的機器學習模型長短期記憶模型(Long Short-Term Memory, LSTM)與空氣品質預測模式。
本研究以廣義可增式模型(Generalized Additive Model, GAM)分析交通因子對於空氣品質的關係，發現車速在 40Km/hr 以下的低車速情況下會加劇初級空氣污染物 CO、NO、NO2、NOX的排放，同時此一現象也反映道路壅塞，車流量大而低車速，造成車流量與空氣污染呈正相關的趨勢。另外本研究也提出新的降雨特徵的分析流程，除了改善傳統降雨分類的不足外，也利用貝氏網路機率圖形模式(Bayesian Network)建立空氣污染的關聯模式，發現降雨因子對於PM2.5 有顯著的影響，該事件內的累積降雨量超過 4mm 即有中等的濕沉降效果，若超過 21mm 則有更顯著的濕沉降效果，而另外本研究也提出新的降雨特徵分類，定義了第Ⅰ型至第Ⅳ型共四類降雨類型其中，代表長延時強降雨的第Ⅰ型降雨有最顯著的濕沉降效果，說明只要降雨事件的延時大於 6 小時都有顯著的濕沉降效果。最後本研究結合 GAM 與貝氏網路模型之分析結果，運用逐步迴歸模型篩選出顯著性因子，並建立 LSTM 模型進行空氣品質的預測模式，預測結果優異，在初級污染物的 CO，NO，NO2，NOX都有約 90%以上的準確率，在次級空氣污染物的 O3，PM2.5，與 PM10 中也有 80%的準確率。
本研究之研究成果量化出車速控制在 40Km/hr 以上，車流量在 1000 輛以內會有最佳的排放模式，並且 6 小時以上降雨或 9mm 以上降雨量會使 PM2.5 下降 20%，可供交通與環境監管單位在政策制訂上的參考。
;There has long shown that air pollution has a serious impact on human health. However, due to the highly non-linear and unstable characteristics of air pollutants, and complex chemical reactions affecting air pollutants from different factors, there is still no a complete air quality analysis process and prediction model. Therefore, with the perspective of data exploration, this research analyzes the actual observational data, and combines traffic factors, atmospheric factors, hydrological conditions, rainfall characteristics and hydrological analysis, and proposes a complete machine learning model (Long Short-Term Memory, LSTM) and air quality prediction model.
This study uses the Generalized Additive Model (GAM) to analyze the relationship between traffic factors and air quality, and finds that the low vehicle speed below 40Km/hr will aggravate the primary air pollutants CO, NO, NO2, and
NOX emissions. This phenomenon also reflects road congestion(high traffic volume and low vehicle speed) resulting in a trend of positive correlation to air pollution. In addition, this research also proposes a new analysis process of rainfall characteristics. This research uses Bayesian Network (BN) to establish an air pollution relationship model. It is found that the rainfall factor has a significant effect on PM2.5. If the cumulative rainfall in the event exceeds 4mm, there will be a moderate wet deposition effect to the PM2.5, and if it exceeds 21mm, there will be a more significant wet deposition effect. In the other hand, Type I rainfall, which represents long-term & heavy rainfall, has the most significant wet deposition effect. The deposition effect shows that the duration of the rainfall event is greater than 6 hours, there will be a
significant wet deposition effect. Finally, this research combines the analysis results of GAM and Bayesian network model, and uses stepwise regression model to filter out non-significant factors to establishes an LSTM model to predict air quality. The prediction results are excellent. The primary pollutants CO, NO, NO2, NOX has an accuracy rate of more than 90%, and also an accuracy rate of 80% in the secondary air pollutants O3, PM2.5, and PM10.
The results of this study quantify that the vehicle speed controlled above 40Km/hr and the vehicle flow less than 1,000 vehicles per hours will have the best
emission mode, and rainfall duration more than 6 hours or accumulative rainfall more than 9mm will reduce 20% of PM2.5 in expectation. The results of this research can provide policy and suggestion to the government. |
