- 學位論文
以支持向量機方法透過時空特徵進行都市空氣品質之預測
Using Support Vector Machine Method with Spatio-temporal Features to Predict Urban Air Quality
摘要
都市空氣品質預測之重要性已被廣泛認可，準確的空氣品質預測可使政府與民眾採取即時的預防、減緩措施。長久以來，空氣品質傳輸模式被大量的研究與應用，然而傳輸模式有其限制，如：準確的排放源資料取得困難、大規模模擬之運算時間較長等，因而具有較為省時、操作容易、調整彈性較大等特性之統計相關預測模型(亦常被稱為機器學習)近年來蓬勃發展。然而大多文獻僅研究時間變異下氣象因子及污染物濃度與空氣品質之相關性，相對較少探討完整時空結合之預測。 本研究以機器學習方法-支持向量機(Support vector machine, SVM)作為基礎，配合地理資訊系統(Geographic information system, GIS)處理空間資料，發展出有效結合時間與空間預測的架構，研究範圍則為北部空品區。本研究首先透過SVM獲取空氣品質指標(Air quality index, AQI)與時間相關特徵(現時空氣品質指標與氣象參數)之訓練模式，並利用此訓練模式完成已知測站的未來空氣品質預測；而後利用此預測值與空間相關特徵(土地利用、交通路網、人口、經濟活動、點源相關、地表高程)之關聯性推估未知空間之未來空氣品質。 驗證結果顯示，僅進行時間預測之準確性很高，四個季節未來一小時的均方根誤差(Root mean square error, RMSE)皆小於4，亦能在更遠的未來時間點(如：未來6小時、12小時)獲得足夠準確的結果。而將時間預測之結果輸入至空間推估後，準確率有較為明顯之下降，四季之RMSE約為10~16，主因可能是空間資料並非動態，較難精確呈現空氣品質之空間變異。空間推估中，與空氣品質最為相關之特徵與人口活動息息相關，如：農林用地面積、交通用地面積、居住用地面積、電力使用量、市區道路密度等，而工業相關之特徵則因為北部空品區內工業相對不發達而影響不大。與其他文獻之機器學習預測表現進行比較，本研究之準確性則相對較低，然而差距並不大，且相較其他文獻使用之方法具備了空間預測之能力。 本研究之結果證實了在實務上進行較現行空氣品質預報更細之時空尺度預測乃是可行的，更細尺度的空氣品質預報可以讓決策者有更精確的依據，此架構也提供了後續有關機器學習研究進行空氣品質之時空預測的參考。
並列摘要
Urban air quality prediction has been considered imperative because it allows citizens to properly respond to poor air quality according to the forecasts. Compared to transport models, statistical methods, usually referring to machine learning, have been more and more popular for air quality prediction in this decade owing to their time-saving and easy-to-use characteristics. However, limited to difficulty in data acquisition, combination of temporal and spatial prediction is still inconclusive. This study aims to utilize support vector machine (SVM), a machine learning algorithm, to predict air quality of unknown space and time with temporal and spatial features extracted by Geographic information system (GIS). The Northern Air Basin of Taiwan was selected as study site; 20 monitoring stations were chosen as training stations (also reference stations) while the rest of 5 stations were testing stations, whose data were not involved in training process. Temporal prediction was first executed in the reference stations, and then the predicted air quality index (AQI) were used for spatial inference to obtain the future AQI of unknown locations. The verification revealed high accuracy of future AQI (the next one hour) prediction with low root mean squared error (RMSE) under 4; nonetheless, higher RMSE over 10 was calculated in spatial inference stage. The performance of spatial inference in winter was noticeably better than the performance in other three seasons probably due to the low spatial divergence of air quality in winter. This “spatio-temporal air quality prediction” is found slightly inaccurate in comparison to other machine learning methods demonstrated in other studies by viewing normalized RMSE; nevertheless, this proposed method is able to conduct spatial prediction, while others can only predict air quality temporally. Despite the fact that the spatial inference only own acceptable accuracy and some obstacles still remain, the framework is feasible in practice with the controlled errors. Further application, like better policy making or more delicate forecasts announced by mobile devices, may be realized under this framework.