Temporal variations and prediction of fine particulate matter (PM _2.5 ) concentrations in Ho Chi Minh City using meteorological data and attention-based deep learning model

Abstract This study investigates the temporal variation and prediction of fine particulate matter (PM 2.5 ) concentrations in Ho Chi Minh City (HCM City), a major Vietnamese metropolis with a tropical monsoon climate, from 2018 to 2023. The results indicated no statistically significant differences in the annual average PM 2.5 concentrations throughout the study period, with values ranging from 21.7 to 26.5 μg m −3 . However, concentrations were consistently higher during the dry season (November to April) (mean ± SD: 27.4 ± 7.3 μg m −3 ) compared to the rainy season (May to November) (mean ± SD: 21.5 ± 6.2 μg m −3 ). PM 2.5 concentrations were strongly negatively correlated with meteorological parameters such as rainfall intensity, ambient air temperature, and wind speed, suggesting removal via wet deposition and enhanced dispersion under stronger winds and higher air temperatures. A bidirectional long short-term memory network with an attention mechanism (BiLSTM+Attention) was proposed to predict PM 2.5 concentrations, incorporating auxiliary variables such as meteorological parameters and the leaf area index from the most recent preceding hours. The model’s best performance was achieved when including both auxiliary variables and PM 2.5 concentrations from the previous 24 h, yielding a coefficient of determination (R 2 ) of 0.944, a mean absolute error of 2.142 μg m −3 , and a root mean square error of 2.957 μg m −3 . Multi-horizon forecasting was also conducted to evaluate the model’s applicability, revealing a decline in prediction accuracy as the forecast horizon increased. SHAP (SHapley Additive exPlanations) was employed to evaluate the contribution of input variables to the model’s outputs, showing that PM 2.5 concentrations from prior hours (e.g, less than 4 h) were the most influential predictors. This study offers new insights into PM 2.5 pollution in HCM City and highlights the potential of advanced deep learning techniques for air quality prediction in tropical monsoon urban environments.