Modelling daily fine particulate matter (PM2.5) and wildfire effects using machine learning and deep learning in a cold urban region

Long-term exposure to PM 2.5 is a major health threat. Therefore, forecasting wildfire-related PM 2.5 plays a pivotal role in informing environmental risks. In this study, three data-driven models, including random forest (RF), long short-term memory (LSTM), and eXtreme Gradient Boosting (XGBoost), were developed to predict the daily PM 2.5 in the Edmonton metropolitan area, Alberta, Canada. Multivariate regression was utilized to correlate PM 2.5 with wildfire events and climate factors. Model performance was evaluated using the coefficient of determination ( R 2 ), Pearson correlation coefficient ( r ), and the root mean square error (RMSE). The results indicate that all three models generally reproduce the variations and patterns of PM2.5 concentrations. However, the models underestimated some unusually high PM 2 . 5 peaks associated with wildfire events. During training, XGBoost across the three stations achieved the highest performance (R 2 = 0.99; RMSE = 0.49 μg/m 3 ), followed by RF (R 2 = 0.87; RMSE = 3.59 μg/m 3 ), whereas LSTM showed substantially lower accuracy (R 2 = 0.30; RMSE = 8.29 μg/m 3 ). In testing, performance declined for all models, with XGBoost (R 2 = 0.86; RMSE = 6.53 μg/m 3 ) and RF (R 2 = 0.75; RMSE = 8.65 μg/m 3 ) remaining comparable, while LSTM continued to underperform (R 2 = 0.27; RMSE = 13.47 μg/m 3 ). Furthermore, Multivariate regression analysis indicates that major PM 2 . 5 peaks (e.g., >100–400 μg/m 3 during May–July 2019 and 2023) coincide with periods of elevated daily burned area, highlighting the strong influence of large wildfires on extreme pollution events. Therefore, ML/DL is a powerful tool for PM 2.5 prediction.