Abstract
This study proposes a gradient-boosting-based machine learning approach for predicting PM<sub>2.5</sub> concentrations in Taiwan. The proposed mechanism is evaluated on a large-scale database built by the Environmental Protection Administration (EPA) and the Central Weather Bureau (CWB), Taiwan, which includes data from 77 air monitoring stations and 580 weather stations performing hourly measurements over one year. By learning from past records of PM<sub>2.5</sub> and neighboring weather stations' climatic information, the forecasting model provides reliable 24-hour predictions. This study also investigates the geographical and meteorological divergence across seven regional monitoring areas. Results indicate that incorporating industrial pollution features from Continuous Emission Monitoring Systems (CEMS) significantly improves prediction accuracy in industrial hubs like Taichung, raising the coefficient of determination (R<sup>2</sup>) from 0.58 to 0.71.
Introduction
Predicting air quality—specifically fine particulate matter (PM<sub>2.5</sub>)—is notoriously difficult in regions with complex geography. Taiwan faces a unique set of challenges: a massive central mountain range that creates stagnant air zones, seasonal monsoons that carry transboundary pollutants, and heavy local industrial clusters. While physical atmospheric models exist, they are computationally expensive and often struggle to account for the "memory effect" of pollutants or the sudden impact of industrial emissions. This paper bridges that gap by using a data-driven machine learning approach that handles high-dimensional, heterogeneous data to provide reliable 24-hour forecasts.
Methods and Datasets
The researchers employed a Gradient Boosting Machine (GBM) based on decision trees. Unlike simple regression, GBM builds an ensemble of shallow trees sequentially, where each new tree attempts to correct the errors (residuals) of the previous ones. The model utilizes an 81-dimensional feature vector including temporal data (weekends/holidays), meteorological data (pressure, temperature, wind vectors), and historical PM<sub>2.5</sub> levels.
The study utilized a massive database from 2017:
- Taiwan EPA: Hourly data from 77 air monitoring stations.
- Central Weather Bureau (CWB): Data from 580 weather stations across the island.
- CEMS Database: Industrial emission records from the Taichung (coal-fired) and Thong-Siau (gas-fired) power plants.
Regional Analysis and Topographic Influence
Taiwan's topography plays a critical role in air quality. The Central Mountain Range (CMR) acts as a barrier to the Northeasterly Monsoon, creating a "lee side" effect in southern Taiwan. This leads to stagnant winds and a massive accumulation of pollutants in regions like Kao-Ping (KP) and Yun-Chia-Nan (YCN).
One of the study's key findings was the similarity between Taipei and London. Both are located in basins, serve as financial centers, and lack heavy domestic industry. Consequently, the model achieved nearly identical performance in both cities (RMSE ~6, R<sup>2</sup> ~0.75), suggesting that urban topography can be a stronger predictor of model performance than geographic proximity.
Key Results
- Industrial Fusion: By adding CEMS features (real-time opacity, flow rates, and gas concentrations), the model's R<sup>2</sup> in Taichung City jumped from 0.58 to 0.71, and the Root-Mean-Square Error (RMSE) dropped from 8.56 to 7.06. This proves that treating power plant emissions as a dynamic, real-time feature is essential for accurate forecasts in industrial zones.
- Regional Accuracy: The model performed exceptionally well in "cleaner" regions like Yilan (YI) and Hua-Dong (HD), where R<sup>2</sup> values approached 0.9, indicating near-perfect tracking of air quality trends.
- Local Variance: Within the Kao-Ping region, Station 68 (Hengchuen, a national park) showed very low error, while Station 69 (Chaozhou, near an industrial area) required the enhanced CEMS model to achieve usable accuracy.
Conclusion
While the GBM approach proved highly effective, predicting PM<sub>2.5</sub> in central Taiwan remains challenging due to the density of pollution sources. Future work will explore deep learning architectures like LSTMs to better capture long-term temporal dependencies and expand CEMS integration to include smaller factories and mobile traffic sources to further refine spatial correlation and prediction accuracy.
Summary updated