Hybrid Al-Driven Framework for Forecasting and Attribution of Urban Air Pollution Episodes: A Decadal Study in Isfahan, Iran

Forecasting acute air pollution episodes, particularly in topographically complex urban basins, remains a critical challenge for environmental management. This study presents and validates a novel hybrid artificial intelligence (AI) framework designed not only to predict but also to attribute severe pollution events in Isfahan, Iran. We integrated a decadal dataset (۲۰۱۴-۲۰۲۳) comprising multi-pollutant concentrations, standard meteorological observations, and derived atmospheric stability indices. The core methodology employs a three-stage sequential AI pipeline. First, a Long Short-Term Memory (LSTM) recurrent neural network was trained to perform ۷۲-hour probabilistic forecasts of PM۲.۵ and AQI. Second, a Gradient Boosting Machine (XGBoost) model was used to conduct a non-linear, non-parametric feature importance analysis, quantifying the conditional contribution of each meteorological driver under different synoptic regimes. Third, an unsupervised learning step (Self-Organizing Maps - SOMS) clustered the pre-episode atmospheric profiles to identify distinct precursor patterns leading to "Unhealthy" air quality. Results demonstrate that the LSTM model achieved a ۹۲% precision in predicting escalation to "Unhealthy" levels ۴۸ hours in advance. The XGBoost attribution revealed that under stagnant high-pressure systems, the combination of planetary boundary layer height (PBLH) suppression and low-level moisture advection were disproportionately significant, often outweighing the role of local wind speed. The SOM analysis identified three dominant precursor clusters, linking specific upper-air trough/ridge configurations to subsequent surface-level pollution accumulation. This hybrid framework moves beyond traditional regression, offering a translatable tool for dynamic, cause-informed air quality forecasting and targeted emission control activation.

کلیدواژه ها:

Air Pollution Forecasting ، Explainable AI (XAI) ، Long Short-Term Memory (LSTM) ، Feature Attribution ، Self-Organizing Maps ، Pollution Episodic Analysis