Melbourne, Australia, faces considerable air quality challenges arising from both urban pollution and seasonal bushfires, which present notable risks to public health. Existing early warning systems often exhibit limitations in predictive accuracy and in providing actionable guidance for communities. This paper introduces SmokeNet, a Transformer-based model that serves as the predictive component of a proposed Proactive Air Quality Forecasting and Health Alert System. The study conducts a comparative evaluation of SmokeNet’s performance against several baselines, including XGBoost and standard recurrent neural networks (LSTM and RNN), using official sensor datasets. The findings indicate superior performance on the overall test set, with SmokeNet achieving a 35.9% reduction in MAE compared to the robust XGBoost baseline. Crucially, this robustness was confirmed in a real-world backtest of a severe smoke event, where it outperformed the same baseline by a substantial 57.7% in MAE. Furthermore, the integration of Explainable AI (XAI) methods with the predictive framework enables the generation of forecasts that are both accurate and interpretable. The results suggest that this approach provides a viable foundation for developing next-generation environmental health services that aim to improve community resilience against air pollution events.