Deep Learning–Based Air Quality Forecasting: A Comparative And Explainable Approach For Multi-Step Pm2.5 Prediction

Abstract Fine particulate matter (PM2.5) is among the most detrimental air pollutants in urban India and precise short-term predictions enable both officials and the public to take pre emptive measures prior to pollution surges. Conventional statistical models have challenges with the non-linear and non-stationary dynamics of air pollution, while solely physical chemical-transport models incur high processing costs. This manuscript introduces a data centric model for predicting PM2.5 levels across many steps (1 to 24 hours in advance) using deep learning techniques. We outline a comprehensive pipeline that encompasses data collection from ongoing ambient air quality monitoring stations, preprocessing, the engineering of meteorological and temporal features, and the training of Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), hybrid CNN-LSTM, and Transformer models in comparison to persistence, ARIMA, and Random Forest benchmarks. Model efficacy is evaluated using RMSE, MAE, and R² via a temporal train-validation-test division, while SHAP is used to elucidate the inputs influencing the predictions. The paper further addresses problems related to data quality, factors to consider during implementation, and inherent constraints. The suggested methodology is replicable, relies only on available data, and may be modified for any monitoring station or urban area.