Abstract
Air quality prediction is a challenging forecasting task due to its spatio-temporal complexity and inherent dynamics. While current models utilize Graph Neural Networks or physics principles, balancing deterministic certainty with stochastic uncertainty remains difficult. We propose Double-Diffusion, a novel diffusion probabilistic model that harnesses known physics to guide air quality forecasting. By integrating a physics-based conditional generative approach with a sampling strategy adapted from image restoration and a new denoiser architecture, Double-Diffusion achieves state-of-the-art results. It reduces inference time by 30% to 50% while improving the Continuous Ranked Probabilistic Score (CRPS) by 3-12% across real-world datasets.
Introduction
Predicting urban air quality, such as PM2.5 levels, is notoriously difficult due to two primary factors: spatio-temporal complexity (pollutant movement across urban grids) and inherent uncertainty (unpredictable shifts in traffic or weather).
Existing solutions typically fall into two categories. Deterministic models, like Physics-Informed Neural Networks (PINNs), capture patterns well but struggle with uncertainty quantification. Conversely, probabilistic models like Denoising Diffusion Probabilistic Models (DDPMs) excel at uncertainty but are computationally expensive and often ignore established physical laws. Double-Diffusion bridges this gap by using physics as a "shortcut" to guide the generative process, making it both more accurate and significantly faster.
Methodology
The "Double-Diffusion" framework refers to the synergy between physical diffusion (pollutant spreading) and probabilistic diffusion (the generative model).
1. Physics-Based Preliminary Forecast
Instead of initializing from pure noise, the model employs a Pollutant Transport Diffusion ODE based on Fick’s Law. This parameter-free step uses historical sensor data and the city's spatial graph to generate a rough physical estimate of pollutant spread.
2. Residual Learning (Resfusion)
Borrowing from image restoration, the model treats the physical forecast as a "degraded" signal. During training, it learns to predict "residual noise"—the specific difference between the physical ODE's output and the ground truth. This allows the model to focus specifically on real-world deviations that simple physics cannot account for.
3. The Zigzag Denoiser
To process these complex sequences, the authors introduced the Zigzag Denoiser. This architecture features:
- Temporal Scaling: It upscales temporal resolution to capture fine-grained features before downscaling, unlike standard U-Nets.
- gMLP Layers: By utilizing Spatial Gating Units (gMLPs) instead of heavy Transformers, the model captures long-range dependencies with significantly lower memory and compute overhead.
Evaluation and Results
The model was tested on two major datasets: Beijing (KDD Cup 2018) and Guangzhou (UrbanAir), predicting the next 12 hours of PM2.5 levels based on a 12-hour history.
- Efficiency: Double-Diffusion is 37% faster than DiffSTG and 53% faster than CSDI. By starting closer to the ground truth via physics guidance, it can skip redundant early diffusion steps.
- Accuracy: The model ranked first in RMSE and CRPS across both datasets. The 3-12% improvement in CRPS indicates that its probability distributions are more reliable for real-world decision-making.
- Robustness: Case studies showed the model successfully captured sudden pollution spikes that pure ODE-based models missed, demonstrating the power of its hybrid approach.
Conclusion
Double-Diffusion demonstrates that physics should not just be a constraint but can serve as a powerful condition for generative models. While currently limited by a simplified diffusion-only ODE and a focus on PM2.5 readings, future iterations could incorporate advection (wind) and auxiliary data like traffic volume to further refine urban air quality forecasting.
Summary updated