HighAir: A Hierarchical Graph Neural Network-Based Air Quality Forecasting Method

Technical Deep Dive: HighAir

Overview

HighAir is a hierarchical graph neural network-based method for forecasting air quality. It constructs a city-level graph and station-level graphs to capture spatial dependencies at multiple granularities. HighAir uses an encoder-decoder architecture, with specialized "upper delivery" and "lower updating" strategies to pass information between the city-level and station-level representations. It also dynamically adjusts edge weights based on wind direction to model the impact of wind on pollutant diffusion.

Methodology

• HighAir constructs a city-level graph and station-level graphs, representing the spatial relationships between cities and monitoring stations respectively.
• It uses a message passing mechanism to model intra-level interactions within the city-level and station-level graphs.
• The "upper delivery" strategy passes historical air quality information from the station-level graphs up to the city-level graph.
• The "lower updating" strategy passes information from the city-level graph down to update the global attributes used in the station-level graphs.
• HighAir uses an encoder-decoder architecture, with the encoder LSTM taking in the station-level node attribute sequences and the decoder LSTM using weather data to generate the final air quality forecasts.

Data & Experimental Setup

• The experiments were conducted on a dataset covering the Yangtze River Delta region in China, including 10 major cities and 61 monitoring stations.
• The dataset contains 1-hour granularity air quality data, POI data, and weather data from January to December 2018.
• The authors evaluated 1-hour, 3-hour, 6-hour, and 12-hour ahead forecasting, using mean absolute error (MAE) and root mean squared error (RMSE) as the evaluation metrics.

Results

• HighAir outperformed various baseline methods, including physical models, time series models, and other graph neural network approaches.
• Ablation studies showed the importance of the hierarchical structure, the city-level LSTM, and the dynamic edge weight adjustment based on wind direction.
• A case study demonstrated HighAir's ability to capture the impact of air pollutant diffusion from neighboring cities, leading to more accurate forecasts during periods of air quality deterioration.

Limitations & Uncertainties

• The authors note that the same graph embedding method has different effects on short-term versus long-term forecasting, suggesting the need for more sophisticated strategies to handle different forecasting horizons.
• They also highlight that the actual pollutant diffusion processes are more complex than the message passing on graphs, and suggest exploring semi-supervised learning to leverage information from regions without monitoring stations.
• Scaling HighAir to larger graph sizes may require techniques like graph pooling to handle over-smoothing issues in deep graph neural networks.

What Comes Next

• Investigating attention mechanisms and meta-learning to automatically learn the importance of topological information for different forecasting horizons.
• Leveraging semi-supervised learning to utilize information from regions without monitoring stations and better model the complex pollutant diffusion processes.
• Validating HighAir's performance on larger-scale datasets and exploring graph pooling methods to handle over-smoothing in deep graph neural networks.