[Paper Review] HighAir: A Hierarchical Graph Neural Network-Based Air Quality Forecasting Method

Ling Chen, Jiahui Xu+5 more|arXiv (Cornell University)|Jan 12, 2021

Air Quality Monitoring and ForecastingEnvironmental Science35 references22 citations

TL;DR

HighAir proposes a hierarchical graph neural network for air quality forecasting that models city-level and station-level pollution patterns using inter-level interactions via upper delivery and lower updating strategies. It dynamically adjusts edge weights based on wind direction and outperforms state-of-the-art methods on the Yangtze River Delta dataset, demonstrating superior accuracy in capturing complex pollutant diffusion and influencing factors like weather and land use.

ABSTRACT

Accurately forecasting air quality is critical to protecting general public from lung and heart diseases. This is a challenging task due to the complicated interactions among distinct pollution sources and various other influencing factors. Existing air quality forecasting methods cannot effectively model the diffusion processes of air pollutants between cities and monitoring stations, which may suddenly deteriorate the air quality of a region. In this paper, we propose HighAir, i.e., a hierarchical graph neural network-based air quality forecasting method, which adopts an encoder-decoder architecture and considers complex air quality influencing factors, e.g., weather and land usage. Specifically, we construct a city-level graph and station-level graphs from a hierarchical perspective, which can consider city-level and station-level patterns, respectively. We design two strategies, i.e., upper delivery and lower updating, to implement the inter-level interactions, and introduce message passing mechanism to implement the intra-level interactions. We dynamically adjust edge weights based on wind direction to model the correlations between dynamic factors and air quality. We compare HighAir with the state-of-the-art air quality forecasting methods on the dataset of Yangtze River Delta city group, which covers 10 major cities within 61,500 km2. The experimental results show that HighAir significantly outperforms other methods.

Motivation & Objective

• To address the challenge of accurately forecasting air quality in regions with complex pollutant diffusion across cities and monitoring stations.
• To model the dynamic interactions between air pollutants, weather conditions, and land usage across spatial hierarchies.
• To improve forecasting performance by capturing both city-level and station-level patterns through a structured hierarchical graph architecture.
• To develop a message-passing mechanism that adapts to changing environmental factors such as wind direction.
• To outperform existing methods in forecasting accuracy by integrating multi-level spatial dependencies and dynamic correlations.

Proposed method

• HighAir employs an encoder-decoder architecture built on a hierarchical graph structure with city-level and station-level subgraphs.
• It introduces two inter-level strategies: upper delivery (aggregating city-level features to stations) and lower updating (refining city-level features using station-level outputs).
• Intra-level interactions are modeled using a message-passing mechanism within each graph level to propagate information among nodes.
• Edge weights in the graphs are dynamically adjusted based on wind direction to reflect the directional transport of pollutants between locations.
• The model incorporates external factors such as weather and land use as node features to enhance predictive accuracy.
• The framework is trained end-to-end on historical air quality, meteorological, and geographical data from 10 cities in the Yangtze River Delta region.

Experimental results

Research questions

• RQ1 How can a hierarchical graph neural network effectively model multi-scale spatial dependencies in air quality forecasting?
• RQ2 To what extent can dynamic edge weighting based on wind direction improve the modeling of pollutant transport between cities and stations?
• RQ3 Can inter-level interactions between city-level and station-level graphs enhance forecasting performance compared to flat or single-level GNNs?
• RQ4 How do external factors like weather and land use influence the predictive accuracy of a graph-based air quality forecasting model?
• RQ5 Does the proposed method outperform existing state-of-the-art approaches on real-world air quality datasets?

Key findings

• HighAir significantly outperforms existing state-of-the-art air quality forecasting methods on the Yangtze River Delta dataset, achieving higher accuracy in predicting PM2.5 and other pollutant concentrations.
• The integration of dynamic edge weights based on wind direction improves the model's ability to capture directional pollutant transport, especially during pollution events.
• The hierarchical design enables the model to learn both macro-level city patterns and micro-level station-level variations, enhancing overall robustness.
• The upper delivery and lower updating strategies effectively facilitate information flow between city and station levels, improving feature representation.
• The message-passing mechanism within each graph level successfully captures local spatial dependencies, contributing to better forecasting performance.
• Empirical results on a 61,500 km² region with 10 major cities confirm that HighAir achieves superior performance across multiple evaluation metrics compared to baseline models.