Deep transfer learning and attention based P2.5 forecasting in Delhi using a decade of winter season data

This investigation focuses on the phenomenon of air pollution in the metropolitan area of Delhi, with a particular emphasis on the stubble-burning season, during which concentrations of PM_2.5 reach their peak, presenting significant health hazards. Utilizing a comprehensive dataset spanning a decade (2012-2022), this study analyzes the influence of meteorological conditions, urban emissions, and seasonal biomass combustion. It amalgamates historical PM_2.5 concentration data, relevant meteorological variables, and FIRECOUNT data to capture the temporal and pollution dynamics. Feature selection based on CorrXGBoost was utilized to find and keep the most significant predictors, hence decreasing model complexity while maintaining predictive efficacy. The proposed hybrid TL-LSTM-MHA Long Short-Term Memory (LSTM) model, augmented with Multi-Head Attention, is employed, harnessing transfer learning techniques to facilitate enhanced computational efficiency and generalization capabilities. The model demonstrated good performance (MAE = 4.38, RMSE = 5.80, R² = 0.9972) and was extensively verified using tenfold cross-validation to ensure robustness towards overfitting and non-stationary effects. Statistical significance tests, particularly the Wilcoxon signed-rank test, were used to confirm the performance disparities among model variations, therefore substantiating the roles of essential architectural elements. Attention weight visualization and head-wise interpretability studies demonstrated unique patterns in feature significance across heads. The model's efficacy was also assessed against traditional and contemporary state-of-the-art methods tested on similar PM_2.5 forecasting tasks, demonstrating its enhanced accuracy. This research provides predictive insights pertinent to regulatory decision-making about seasonal air quality management encountered in Delhi. The scalability of the proposed framework is demonstrated by comparing it to conventional and transfer learning-based models.