Accurate and interpretable air pollution forecasting is essential for public health and proactive intervention. While deep learning models have achieved strong performance in long-horizon forecasting, they often require aligned historical sequences and provide limited interpretability. To address these limitations, we propose a novel Transformer-based framework that integrates historical observations with auxiliary forecasts to enhance predictive performance while providing built-in interpretability. The framework introduces two core components: a Structured Temporal Embedding Module (STEM) for preserving the semantic integrity of heterogeneous inputs, and an eXplainable Target-Oriented Cross-Attention (X2-Attention) mechanism that aligns future predictions with known inputs to yield fine-grained attribution for each forecasted timestep across both temporal and feature dimensions. Evaluated on diverse urban and suburban datasets in Stockholm, our model simultaneously delivers reliable interpretability and superior predictive accuracy. Compared to five baselines (TimeXer, Crossformer, iTransformer, Transformer and LSTM), it exhibits significantly lower Mean Squared Error (MSE) for NOX and competitive performance for PM10 across horizons up to 720 hours. Ablation studies reveal that incorporating auxiliary forecasts reduces MSE by 29 % for NOX and 15 % for PM10. Moreover, the X2-Attention mechanism generates plausible feature attribution validated through perturbation analysis. Its attributions are consistent with GradientSHAP while avoiding prohibitive computational overhead, thereby facilitating efficient air quality forecasting.

As air pollution is regarded as the single largest environmental health risk in Europe it is important that communication to the public is up to date and accurate and provides means to avoid exposure to high air pollution levels. Long- and short-term exposure to outdoor air pollution is associated with increased risks of mortality and morbidity. Up-to-date information on present and coming days' air quality helps people avoid exposure during episodes with high levels of air pollution. Air quality forecasts can be based on deterministic dispersion modelling, but to be accurate this requires detailed information on future emissions, meteorological conditions and process-oriented dispersion modelling. In this paper, we apply different machine learning (ML) algorithms - random forest (RF), extreme gradient boosting (XGB), and long short-term memory (LSTM) - to improve 1, 2, and 3d deterministic forecasts of PM10, NOx, and O3 at different sites in Greater Stockholm, Sweden. It is shown that the deterministic forecasts can be significantly improved using the ML models but that the degree of improvement of the deterministic forecasts depends more on pollutant and site than on what ML algorithm is applied. Also, four feature importance methods, namely the mean decrease in impurity (MDI) method, permutation method, gradient-based method, and Shapley additive explanations (SHAP) method, are utilized to identify significant features that are common and robust across all models and methods for a pollutant. Deterministic forecasts of PM10 are improved by the ML models through the input of lagged measurements and Julian day partly reflecting seasonal variations not properly parameterized in the deterministic forecasts. A systematic discrepancy by the deterministic forecasts in the diurnal cycle of NOx is removed by the ML models considering lagged measurements and calendar data like hour and weekday, reflecting the influence of local traffic emissions. For O3 at the urban background site, the local photochemistry is not properly accounted for by the relatively coarse Copernicus Atmosphere Monitoring Service ensemble model (CAMS) used here for forecasting O3 but is compensated for using the ML models by taking lagged measurements into account. Through multiple repetitions of the training process, the resulting ML models achieved improvements for all sites and pollutants. For NOx at street canyon sites, mean squared error (MSE) decreased by up to 60%, and seven metrics, such as R2 and mean absolute percentage error (MAPE), exhibited consistent results. The prediction of PM10 is improved significantly at the urban background site, whereas the ML models at street sites have difficulty capturing more information. The prediction accuracy of O3 also modestly increased, with differences between metrics. Further work is needed to reduce deviations between model results and measurements for short periods with relatively high concentrations (peaks) at the street canyon sites. Such peaks can be due to a combination of non-typical emissions and unfavourable meteorological conditions, which are rather difficult to forecast. Furthermore, we show that general models trained using data from selected street sites can improve the deterministic forecasts of NOx at the station not involved in model training. For PM10 this was only possible using more complex LSTM models. An important aspect to consider when choosing ML algorithms is the computational requirements for training the models in the deployment of the system. Tree-based models (RF and XGB) require fewer computational resources and yield comparable performance in comparison to LSTM. Therefore, tree-based models are now implemented operationally in the forecasts of air pollution and health risks in Stockholm. Nevertheless, there is big potential to develop generic models using advanced ML to take into account not only local temporal variation but also spatial variation at different stations.

Forecasting air pollution is an important activity for developing sustainable and smart cities. Generated by various sources, air pollutants distribute in the atmospheric environment due to the complex dispersion processes. The emerging sensor and data technologies have promoted the development of data-driven approaches to replace conventional physical models in urban air pollution forecasting. Nevertheless, it is still challenging to capture the intricate spatial and temporal patterns of air pollutant concentrations measured by heterogeneous sensors, especially for long-term prediction of the multi-variate time series data. This paper proposes a deep learning framework for longer-term forecast of air pollutants concentrations using air pollution sensing data, based on a conceptual framework of meta-graph deep learning. The key modules in the framework include meta-graph units and fusion layers, which are designed to learn temporal and spatial correlations respectively. A detailed case was formulated for forecasting air pollutants in Stockholm using air quality sensing data, meteorological data and so on. Experiments were conducted to evaluate the performance of the proposed modelling framework. The computational results show that it outperforms the baseline models and conventional deterministic dispersion models, demonstrating the potential of the framework to be deployed for the real air quality information systems in Stockholm.