Multi-horizon Forecasting of Air Pollutants Using an Explainable Multi-task Attention-Based LSTM

Precise prediction of air pollutants is of crucial importance in urban environmental management and public health. In this article, we propose a multi-task and multi-horizon deep learning framework to forecast nitrogen dioxide (NO2) and ozone (O3) concentrations up to 12 h ahead using multivariate time series data. The model, based on an attention-enhanced Long Short-Term Memory (LSTM) network, learns temporal interactions between meteorological and pollutant variables from different monitoring stations. Multi-horizon forecasting is supported by a common architecture that makes simultaneous predictions of NO2 and O3 for 1, 3, 6, and 12-hour forecasting horizons. Optuna is used to tune hyperparameters, and model explainability is achieved through integrated gradients and attention visualization. Experimental findings demonstrate that the proposed model achieved high predictive performance, with explainability tools offering valuable understanding into temporal and feature-level importance. This approach shows the potential of explainable deep learning for high-fidelity air quality forecasting in urban areas.