물리 기반 해석 가능한 시공간 대기오염 예측 프레임워크

📝 Abstract

Accurate and interpretable air pollution forecasting is crucial for public health, but most models face a trade-off between performance and interpretability. This study proposes a physics-guided, interpretable-by-design spatiotemporal learning framework. The model decomposes the spatiotemporal behavior of air pollutant concentrations into two transparent, additive modules. The first is a physics-guided transport kernel with directed weights conditioned on wind and geography (advection). The second is an explainable attention mechanism that learns local responses and attributes future concentrations to specific historical lags and exogenous drivers. Evaluated on a comprehensive dataset from the Stockholm region, our model consistently outperforms state-of-the-art baselines across multiple forecasting horizons. Our model’s integration of high predictive performance and spatiotemporal interpretability provides a more reliable foundation for operational air-quality management in real-world applications.

💡 Analysis

Accurate and interpretable air pollution forecasting is crucial for public health, but most models face a trade-off between performance and interpretability. This study proposes a physics-guided, interpretable-by-design spatiotemporal learning framework. The model decomposes the spatiotemporal behavior of air pollutant concentrations into two transparent, additive modules. The first is a physics-guided transport kernel with directed weights conditioned on wind and geography (advection). The second is an explainable attention mechanism that learns local responses and attributes future concentrations to specific historical lags and exogenous drivers. Evaluated on a comprehensive dataset from the Stockholm region, our model consistently outperforms state-of-the-art baselines across multiple forecasting horizons. Our model’s integration of high predictive performance and spatiotemporal interpretability provides a more reliable foundation for operational air-quality management in real-world applications.

📄 Content

Accurate and reliable urban air pollution forecasts are vital for public health and environmental management. They enable municipal authorities to issue timely health warnings and implement evidence-based traffic management [1]. The evolution of pollutant concentrations across citywide monitoring networks can be described by the advection-diffusionreaction equation [2]. Two processes dominate: i) Spatial transport, governed mainly by advection along the wind and by near-surface mixing, and ii) Local processes, arising from site-specific emissions and chemical transformation. Existing modeling paradigms struggle to effectively represent and disentangle these processes, leading to a persistent tradeoff between predictive accuracy and model interpretability [3], [4].

Prevailing approaches include numerical and data-driven models. Numerical models provide physical fidelity but are constrained by heavy computation and high-resolution inventories, limiting real-time applications [5]. Deep learning models scale well and achieve strong accuracy [6]- [8], yet they compress physical structures into opaque representations with limited diagnostic value. While hybrid physical-neural network designs show promise for improving accuracy [9], [10], they often couple training to costly numerical simulators and their attributions remain opaque.

We propose a physics-guided spatiotemporal learning framework that is interpretable-by-design. Our model aligns with physical principles by decoupling the pollutant forecasting into two additive modules with transparent attribution, illustrated in Fig 1 (b) and(c). The first module captures station-wise temporal dependencies using a variant of interpretable attention mechanism. A learnable query attends over historical pollutant and auxiliary-variable sequences, attributing the forecast to specific lags and exogenous variables [3]. The second module learns a physics-guided kernel for spatial transport. Its directed weights are dynamically conditioned on meteorological fields and geographic relationships, forming a learnable spatial operator for advection. This separation provides clear physical meaning to each component and enhances both predictive accuracy and interpretability.

The main contributions are summarized as follows:

• An interpretable spatiotemporal framework that separates spatial transport from local dynamics and aligns with physical principles; • A physics-guided, time-varying kernel capturing crossstation interactions underlying the advection process; • A variant of interpretable attention mechanism providing station-level attributions that map each input variable to its contribution at each forecast step; • Extensive experiments on a city-scale case study demonstrate superior accuracy over state-of-the-art baselines and yield spatiotemporal interpretations validated through a case analysis.

Numerical air quality models, spanning from regional Chemical Transport Models (CTMs) [11] to local street-level models [12], provide physical grounding. However, they are computationally intensive and highly sensitive to the accuracy of emission inventories and meteorological inputs, limiting their operational agility in complex urban environments [8]. Data-driven methods, including Recurrent Neural Networks (RNNs), Convolutional Neural Networks (CNNs), Graph Neural Networks (GNNs), and Transformers, excel at modeling complex nonlinear dependencies [7], [13], [14]. A persistent challenge, however, is interpretability. Many GNNs employ static, distance-based graphs, which are a poor surrogate for wind-driven atmospheric transport [6]. More advanced dynamic graph models often derive connectivity from statistical correlations, resulting in opaque representations that are disconnected from underlying physical processes like advection.

To build more trustworthy models, researchers are integrating physical knowledge into neural networks. One approach, Physics-Informed Neural Networks (PINNs), uses governing PDEs as soft constraints [15]. This guides the model toward physically plausible solutions but does not render the network architecture inherently interpretable. Another direction involves hybrid models that use neural networks to parameterize components of a physical process, often formulated with Ordinary Differential Equations (ODEs). For instance, AirPhyNet uses a GNN-based differential equation network to model the physical transport of air particles, specifically diffusion and advection [9]. While these methods improve physical consistency, their reliance on iterative ODE solvers can introduce significant computational overhead during training. Furthermore, their intricate coupling of neural and physical components can still obscure clear attribution.

We forecast PM 10 at a network of S stations over a future horizon of H hours, conditioned on a look-back window of L hours. All series are sampled hourly.

Features are indexed by i ∈ {1, . . . , F } and stations by s ∈ {1,

This content is AI-processed based on ArXiv data.