Date and time: Friday 7 November 2025, 10:00-11:30 CET
Speaker: Zhiguo Zhang, Doctoral Student, Transport Planning, BYV, KTH
Title: Interpretable Machine Learning Models for Multi-Day Air Pollutant Forecasting
Where: M108, Brinellvägen 23, KTH main Campus
Discussant: Frida Green, Senior Scientist in ML, SCB Sverige
Abstract: Accurate and interpretable air pollution forecasting is essential for effective public health management, enabling timely alerts and interventions. Existing methods face a critical trade-off between forecasting accuracy and interpretability. Physics-based deterministic models provide mechanistic insight but suffer from systematic biases, whereas data-driven models achieve high accuracy yet typically lack transparency. A key challenge is developing frameworks that improve accuracy while providing dynamic interpretability beyond static feature assessments, especially when integrating heterogeneous data sources such as historical observations, meteorological predictions, and deterministic model outputs.
This study explores a progression of machine learning (ML) techniques to overcome these limitations. Initial work confirmed that single-step ML models, including RandomForest, XGBoost, and Long Short-Term Memory, significantly improve upon deterministic forecasts. Building on this foundation, key contributions include interpretable Transformer-based frameworks designed for long-horizon prediction. These frameworks utilize structured embedding modules to effectively integrate heterogeneous, forward-looking inputs and feature a novel attention mechanism (X2-Attention) that explicitly links future targets to relevant inputs, thereby providing built-in, timestep-wise feature attributions. This architecture was extended to multi-target forecasting ($NO_X$ and $PM_{10}$) and integrated into a “Predict-Validate-Interpret-Optimize” workflow. Hierarchical interpretability analysis elucidates model decision-making processes and enables practical model optimization. To address spatial dynamics, a physics-guided spatiotemporal learning approach further advances these methods by decoupling pollutant behavior. It combines a learnable, wind-conditioned advection kernel modeling spatial transport with an attention-based module capturing local responses, yielding accurate spatiotemporal predictions alongside interpretable cross-station attributions.
Extensive evaluations in Stockholm confirm that these advanced ML models consistently outperform deterministic models and advanced baselines across various pollutants and forecast horizons up to 720 hours, underscoring the critical role of integrating auxiliary forecast data for enhancing long-horizon accuracy. Additionally, the interpretable designs successfully elucidate dynamic drivers by revealing shifts in feature importance over time and during local meteorological events. Furthermore, the physics-guided spatial model accurately represents transport phenomena consistent with meteorological conditions. These results collectively establish that high predictive accuracy and robust, dynamic interpretability can be concurrently achieved, offering a viable pathway towards developing more reliable, trustworthy, and actionable air quality forecasting systems.
The work is partially funded by iHorse and iHorse+ through Digital Futures.
