[Paper Review] AirPhyNet: Harnessing Physics-Guided Neural Networks for Air Quality Prediction

Kethmi Hirushini Hettige, Jiahao Ji+4 more|arXiv (Cornell University)|Feb 6, 2024

Air Quality Monitoring and Forecasting5 citations

TL;DR

AirPhyNet integrates diffusion-advection physics into a neural network to predict PM2.5 with improved accuracy, especially in sparse data scenarios, outperforming state-of-the-art baselines.

ABSTRACT

Air quality prediction and modelling plays a pivotal role in public health and environment management, for individuals and authorities to make informed decisions. Although traditional data-driven models have shown promise in this domain, their long-term prediction accuracy can be limited, especially in scenarios with sparse or incomplete data and they often rely on black-box deep learning structures that lack solid physical foundation leading to reduced transparency and interpretability in predictions. To address these limitations, this paper presents a novel approach named Physics guided Neural Network for Air Quality Prediction (AirPhyNet). Specifically, we leverage two well-established physics principles of air particle movement (diffusion and advection) by representing them as differential equation networks. Then, we utilize a graph structure to integrate physics knowledge into a neural network architecture and exploit latent representations to capture spatio-temporal relationships within the air quality data. Experiments on two real-world benchmark datasets demonstrate that AirPhyNet outperforms state-of-the-art models for different testing scenarios including different lead time (24h, 48h, 72h), sparse data and sudden change prediction, achieving reduction in prediction errors up to 10%. Moreover, a case study further validates that our model captures underlying physical processes of particle movement and generates accurate predictions with real physical meaning.

Motivation & Objective

• Motivate accurate outdoor PM2.5 forecasting by blending physics with data-driven learning.
• Incorporate fundamental air transport physics (diffusion and advection) into a deep learning framework.
• Enhance prediction accuracy and interpretability, especially under sparse data and abrupt changes.

Proposed method

• Encode historical PM2.5 and weather data with an RNN encoder (GRU).
• Model diffusion and advection on a sensor network as differential equation networks with graph Laplacians.
• Fuse diffusion and advection dynamics via a gated mechanism to form a Diffusion-Advection Differential Equation (DE).
• Parameterize the DE network with a GNN-informed neural ODE to evolve latent states over time.
• Decode latent predictions to PM2.5 concentrations at future horizons (up to 72 hours).
• Train end-to-end by minimizing MAE between predicted and ground-truth concentrations.

Figure 1: The overall architecture of AirPhyNet framework consisting of a RNN Encoder, GNN based DE Network and a Decoder. Heatmap indicate the PM2.5 concentrations while the dashed arrows represent the air pollutant transport between nodes due to diffusion and advection.

Experimental results

Research questions

• RQ1 Can physics-guided deep learning (diffusion-advection) improve outdoor PM2.5 forecasting over purely data-driven models?
• RQ2 Do physics-informed components enhance performance in sparse data and sudden-change scenarios?
• RQ3 Does the proposed GNN-based differential equation network effectively capture spatiotemporal transport dynamics on sensor networks?
• RQ4 What is the interpretability benefit of linking predictions to underlying diffusion and advection processes?

Key findings

• AirPhyNet consistently outperforms baselines across both Beijing and Shenzhen datasets for 24h, 48h, and 72h forecasts.
• Mean improvements over the second-best method are 3.7% in MAE and 6.1% in RMSE on average in full-data settings.
• In sparse data settings, AirPhyNet yields 8.3% (MAE) and 8.1% (RMSE) improvements on average over competitors.
• The model achieves up to 8% reduction in prediction errors for sudden-change scenarios.
• Ablation studies show advantages of combining diffusion and advection, and of using a GRU encoder over a TCN-based encoder.
• Case study visualizations illustrate physically meaningful diffusion and advection patterns in predictions.

(a) Effect of Physical Knowledge in AirPhyNet