Abstract Data-driven models have exhibited strong predictive accuracy in air quality forecasting, yet their inherent “black-box” nature remains a primary barrier to operational adoption. To address this gap, we develop X2-AQFormer by extending an explainable Transformer-based framework for multi-target, multi-day forecasting of NO X and PM 10 . Through extensive evaluation across multiple street-canyon and urban background sites in Stockholm, Sweden, our model demonstrates substantial performance improvements over forecast horizons up to 72 hours. The framework systematically corrects the biases of deterministic models and outperforms widely used tree-based models and other advanced Transformer-based architectures. Beyond predictive accuracy, the model’s built-in interpretability reveals the dynamic drivers of NO X and PM 10 from multiple perspectives. Specifically, we analyze overall feature importance rankings, the temporal evolution of these drivers, and the model’s dynamic responses to key meteorological events like precipitation. These intrinsic explanatory insights show strong alignment with the outputs of the post-hoc GradientSHAP method. Finally, case studies reveal that using only the top-ranked feature subsets identified through the model’s explanations can maintain near-optimal performance, highlighting the approach’s practical utility. By presenting a “Predict-Validate-Interpret-Optimize” workflow, this study provides a feasible pathway toward developing reliable, trustworthy, and actionable forecasting tools.
Zhang et al. (Mon,) studied this question.