| 摘要: | 許多研究指出，細懸浮微粒（PM2.5）不僅對環境與氣候變遷產生影響，亦對人類健康帶來嚴重威脅。因此，發展具有高準確率的空氣品質預報系統，以提供有效的預警與應變資訊，具有重要意義。臺灣目前的空氣品質預報作業主要仰賴CMAQ （Community Multiscale Air Quality model）空氣品質模式，其所需之氣象參數使用WRF（Weather Research and Forecasting）氣象模式的模擬結果，而排放量資料則來自環境部提供之臺灣空氣污染物排放量清冊（Taiwan Emission Data System, TEDS）。儘管此類基於大氣動力以及污染物傳輸與化學轉化過程的數值模式能有效掌握污染物濃度的變化趨勢，但由於模式各項機制參數化所採用的理想性假設與輸入資料本身存在的不確定性，導致數值預報結果仍存在一定誤差。
隨著人工智慧技術的發展，近年來愈來愈多研究開始導入深度學習演算法，以建構預報能力更佳的空氣品質預報模型。本研究採用卷積神經網路（Convolutional Neural Network, CNN）並搭配一維卷積架構用以擷取濃度隨時間變化特徵，建構具高預報能力之PM2.5濃度預報模型。模型整合了環境部空氣品質監測站之每小時PM2.5濃度觀測值、現行空氣品質預報系統（AQF）的預報結果，以及綜觀天氣型態類別（Synoptic Weather Patterns, SWPs），以預測全臺75個監測站未來72小時內的PM2.5濃度變化。模型訓練資料期間涵蓋2019年10月至2021年9月，並以2021年10月至2022年9月的資料進行驗證與測試，以確保模型具備良好的泛化能力與穩定性。
本研究共建構五種CNN模型，分別為：未納入土地利用型態與SWPs的CNN-noLU模型、未納入SWPs資訊的CNN-BASE模型、結合SWPs資訊之CNN-SWP模型、與CNN-SWP使用相同輸入參數，但進一步採用加權損失函數（weighted loss function）處理PM2.5濃度值數據分布不均問題之CNN-SWPW模型，以及與CNN-BASE使用相同輸入參數，但分空品區建構模型。
評估結果顯示，CNN-BASE可將72小時PM2.5濃度預測的平均根均方誤差（Root Mean Square Error, RMSE）由現行AQF的10.48 μg/m³ 明顯降低至6.88 μg/m³。然而，觀測中屬於高濃度事件（PM2.5日均值 ≥ 35.5 μg/m3）僅占整體樣本的3.4%，導致 CNN-BASE於此濃度範圍的預測準確率僅為26.2%。對比CNN-BASE，CNN-noLU與CNN-AQZ皆具有較高的RMSE與相對誤差（Relative Error, RE），說明地理資訊能提升模型的預報能力且分區建構模型並無太大幫助。
考量臺灣中南部高污染事件常 與不利污染物擴散的氣象條件相關，CNN-SWP藉由加入SWP資訊後，有效提升中、高PM2.5濃度的預報表現。而CNN-SWPW透過加權損失函數強化對少數類別樣本的學習能力，使高濃度事件的預測準確率提升至65.7%。
整體而言，本研究展現了CNN模型應用於空氣品質預報上的潛力，所建構之CNN模型可與現行AQF系統整合，對於提升預報準確性與政策決策支援能力具有重要貢獻。;Many studies have demonstrated that fine particulate matter (PM2.5) adversely impacts the environment, exacerbates climate change, and poses a significant risk to human health. Therefore, developing a highly accurate air quality forecasting system has become an urgent and critical task. In Taiwan, the current operational air quality forecasting system (AQF) is based on the Community Multiscale Air Quality (CMAQ) model, which uses emission data from the Taiwan Emission Data System (TEDS) provided by the Ministry of Environment (MOENV) and meteorological parameters simulated from the Weather Research and Forecasting (WRF) model. Although numerical models grounded in atmospheric dynamics and pollutant governing equations can yield reliable long-term trends in pollutant concentrations, forecasting biases remain due to idealized assumptions in model parameterizations and uncertainties from the input data.
With the advancement of artificial intelligence technologies, an increasing number of recent studies have adopted deep learning algorithms to develop air quality forecasting models with superior predictive capabilities. To extract temporal features of pollutants’ concentrations and construct an accurate forecasting model, this study utilized a Convolutional Neural Network (CNN) with a one-dimensional convolutional operation. The constructed model integrates hourly PM2.5 concentration observed from the air quality monitoring stations provided by the MOENV, simulated from the AQF system, and synoptic weather patterns (SWPs) to forecast PM2.5 concentration over the next 72 hours at 75 monitoring sites across Taiwan. The training dataset spans from October 2019 to September 2021, while the data from October 2021 to September 2022 is used for validation and testing to ensure the generalization of the forecasting model. This study developed five CNN models for PM2.5 forecasting: a baseline model excluding meteorological parameters (CNN-BASE), a model without the land use index (CNN-noLU), a model incorporating synoptic weather patterns (CNN-SWP), a model employing a weighted loss function to mitigate the imbalance in PM2.5 concentrations (CNN-SWPW), and a region-specific model using the same input parameters as CNN-BASE but trained separately for each air quality regions (CNN-AQZ).
The evaluation findings indicate that the CNN-BASE model significantly decreased the root mean square error (RMSE) of the 72-hour PM2.5 forecast from 10.48 μg/m³ of AQF to 6.88 μg/m³. Nevertheless, the restricted representation of high pollution events (daily PM2.5 ≥ 35.5 μg/m³), comprising merely 3.4% of the dataset, resulted in a CNN-BASE accuracy of only 26.2% within this concentration range. In comparison with CNN-BASE, both CNN-noLU and CNN-AQZ demonstrate higher RMSE and relative error (RE), suggesting that the inclusion of geographic information improves forecasting accuracy, while constructing region-specific models provides limited benefit.
Since the high pollution events in Taiwan are often associated with meteorological conditions, the CNN-SWP model showed improved forecasting performance for moderate to high PM2.5 concentrations by incorporating SWPs. The CNN-SWPW model further enhanced the accuracy under high concentration scenarios by strengthening the learning of minority samples through a weighted loss function, achieving an accuracy rate of 65.7% in these challenging situations.
Overall, the results demonstrate the potential of CNN models in operational air quality forecasting and their significant contribution to enhancing the accuracy and decision-support capabilities of existing forecasting systems. |
