Source-linked AI summary
PM2.5-GNN: A Domain Knowledge Enhanced Graph Neural Network For PM2.5 Forecasting
Shuo Wang, Yanran Li, Jiang Zhang, Qingye Meng, Lingwei Meng, Fei Gao
TL;DR
PM2.5 forecasting requires handling domain-sensitive processes and long-term effects. The paper develops PM2.5-GNN, combining a knowledge-enhanced directed graph model with recurrent temporal modeling, and reports strong performance across three datasets alongside 72-hour forecasting deployment.
Problem
PM2.5 prediction must model domain-sensitive transport and diffusion processes whose effects can persist over long periods.
Method
PM2.5-GNN uses a directed city graph with meteorological and geographical knowledge, a knowledge-enhanced GNN for spatial transport, and an RNN for temporal diffusion.
Results
PM2.5-GNN surpasses compared models on almost every metric across three sub-datasets and achieves the best test loss, RMSE, MAE, CSI, and POD on all three datasets.
Takeaways & Limitations
The approach models fine-grained and long-term PM2.5 dependencies and has been deployed online to provide real-time forecasting through free API access.
Abstract
from arXiv · showhide
When predicting PM2.5 concentrations, it is necessary to consider complex information sources since the concentrations are influenced by various factors within a long period. In this paper, we identify a set of critical domain knowledge for PM2.5 forecasting and develop a novel graph based model, PM2.5-GNN, being capable of capturing long-term dependencies. On a real-world dataset, we validate the effectiveness of the proposed model and examine its abilities of capturing both fine-grained and long-term influences in PM2.5 process. The proposed PM2.5-GNN has also been deployed online to provide free forecasting service.
1 INTRODUCTION
PM2.5 forecasting must account for domain knowledge and long-term dependencies because concentrations arise from transport and diffusion shaped by meteorological and geographical factors. PM2.5-GNN addresses these characteristics with a knowledge-enhanced directed graph and recurrent modeling, and supports 72-hour forecasting.
- PM2.5 concentrations reflect a complex process of emission, transport, and diffusion influenced by meteorological and geographical information.
- PM2.5 factors have wide-range and long-lasting effects, with particles potentially transported hundreds of kilometers in 72 hours.
- Existing graph-based approaches model spatial dependencies but do not explicitly incorporate critical domain knowledge such as wind directions.
- PM2.5-GNN uses a directed city graph, meteorological node and edge features, geographical graph structure, a knowledge-enhanced GNN, and an RNN for temporal diffusion.
- The paper introduces the KnowAir dataset, evaluates 72-hour PM2.5 prediction, and deploys the model through a free online forecasting service.
2 RELATED WORK
Prior PM2.5 forecasting methods use image- or graph-based representations, while PM2.5-GNN combines graph modeling with recurrent learning to represent transport and diffusion over long periods. The paper evaluates this approach on the large-scale KnowAir dataset.
- Deep-learning PM2.5 forecasting methods are broadly classified as image-based or graph-based according to their input-data structure.
- Image-based methods interpolate monitoring-station concentrations onto images, but uneven urban and rural station distributions can cause imprecise process modeling.
- Graph-based methods preserve monitoring locations and learn spatial or spatiotemporal dependencies through graph structures, diffusion convolution, or graph convolution.
- GNNs propagate node information along graph edges through message passing and neighborhood aggregation.
- PM2.5-GNN integrates a GNN with an RNN to model transport and diffusion over a long term and tests air forecasting on the large-scale KnowAir dataset.
3 METHODOLOGY
The methodology frames PM2.5 forecasting over a wide geographic area as a spatiotemporal sequence problem and constructs a knowledge-aware directed city graph. PM2.5-GNN combines direction-aware graph message passing for horizontal transport with a spatio-temporal GRU for vertical accumulation and diffusion.
- 3.1 Studied Area and Data: The study uses a wide area in China to examine models’ ability to capture long-term dependencies.The area spans 103°E–122°E and 28°N–42°N and covers several severely polluted regions.
- 3.2 Problem Definition: PM2.5 forecasting is formulated as spatiotemporal sequence prediction over city nodes and directed edges representing potential city interactions.Node and edge attribute matrices describe meteorological and interaction-related information at each time step.
- 3.2 Problem Definition: The model feeds observed PM2.5, next-T-step attribute sequences, and the graph structure into T iterative PM2.5-GNN steps.The initial observed concentration is used when previous predictions are unavailable.
- 3.3 Graph Construction: Node attributes encode meteorological effects including PBL height, K index, wind speed, temperature, humidity, and precipitation.These variables represent mechanisms affecting dilution, instability, transport, chemical formation, and wet scavenging.
- 3.3 Graph Construction: Edge attributes and adjacency structure encode transport using source-sink variables, wind direction, distance, and terrain constraints.Edges permit transport only when cities are within 300 km and intervening mountains are lower than 1200 m.
- 3.4 PM2.5-GNN Model: A knowledge-enhanced GNN captures horizontal pollutant transport, while a spatio-temporal GRU models vertical accumulation and diffusion under weather influence.Iterative message passing aggregates direction-aware neighboring influences before recurrent processing.
4 EXPERIMENTS
Experiments evaluate PM2.5-GNN on KnowAir across multiple forecasting settings, comparing it with temporal, graph-based, and fully connected baselines. The model achieves the strongest reported overall performance and exhibits benefits from neighboring information, domain knowledge, and long-term modeling.
- 4.1 Datasets: KnowAir contains four years of data from 184 cities and is divided into three sub-datasets for examining general, heating-season, and other forecasting settings.Dataset 1 uses a 2:1:1 train/validation/test split; Dataset 2 focuses on November–February heating seasons.
- 4.2 Compared Models: The task predicts the next 72 hours of PM2.5 concentrations from initial observations and weather forecasts, using MLP, LSTM, GRU, GC-LSTM, nodesFC-GRU, and PM2.5-GNN.The compared models differ in their use of memory, graph structure, neighborhood information, and domain knowledge.
- 4.3 Experimental Settings: Evaluation averages 24 city-by-time metrics across 184 cities and repeats experiments 10 times, reporting losses, MAE, RMSE, CSI, POD, and FAR.CSI, POD, and FAR assess performance near the 75 μg/m^3 pollution threshold.
- 4.4 Experimental Results: MLP, LSTM, and GRU perform worst on automatic metrics, supporting the necessity of neighboring information for PM2.5 prediction.These models access only node representations, while the other compared models capture both temporal and spatial dependencies.
- 4.4 Experimental Results: PM2.5-GNN achieves the best test loss, RMSE, MAE, CSI, and POD on all three datasets and surpasses compared models on almost every metric.The reported result indicates stronger exploitation of domain knowledge and reliable prediction ability across the sub-datasets.
- 4.4 Experimental Results: Fine-grained analyses show that memory improves longer-horizon predictions, while fully connected neighborhood modeling can overfit or introduce excessive dependencies in regions with few transport pathways.In East China, GC-LSTM and PM2.5-GNN perform best, whereas nodesFC-GRU performs worst; removing domain-knowledge components also degrades performance substantially.
5 CONCLUSIONS
PM2.5-GNN combines domain knowledge with GNN and RNN components to model fine-grained and long-term dependencies in PM2.5 processes. The approach is evaluated through extensive experiments and deployed online for forecasting services.
- PM2.5-GNN integrates a GNN with an RNN to capture fine-grained and long-term dependencies in the PM2.5 process.
- The model incorporates domain knowledge into graph-structured data for PM2.5 prediction.
- Extensive experiments are used to demonstrate the approach, and PM2.5-GNN is deployed online for forecasting.
A DEPLOYMENT DETAILS
The online deployment serves PM2.5 forecasts through websites, mobile apps, and APIs. It uses online weather and environmental data, updates inputs daily, and retrains the model every midnight.
- PM2.5-GNN is deployed through an air forecasting website, mobile apps, and API protocols for third-party clients.
- Offline experiments use ERA5 reanalysis data, which is unavailable in real time and therefore unsuitable for the online service.
- The online framework stores MEE and GFS raw data in a spatio-temporal database for service use.
- Because PM2.5 forecasting is sensitive to meteorological and related domain knowledge, online inputs are updated daily and the model is retrained every midnight.
B STUDIED AREA
KnowAir covers a broad geographic area in China, including severely polluted regions. Its studied-area graph represents cities and potential PM2.5 transport interactions based on distance and geographic barriers.
- KnowAir covers a wide range of areas for PM2.5 forecasting, including several severely polluted regions in China.The stated coordinate range is 103°E–122°E and 28°N–42°N.
- The studied-area map marks cities as dots and potential inter-city interactions as blue lines.
- A potential interaction is defined when cities are within a distance threshold and no high mountains obstruct PM2.5 transport.
- The KnowAir studied area is substantially larger than those used in previous work.
C PM2.5-GNN PROCEDURE
The PM2.5-GNN procedure iterates over time and graph nodes, applying graph and recurrent operations before producing predictions. The procedure is presented in an algorithmic form alongside model and deployment figures.
- The learning procedure is summarized as Algorithm 1 for the PM2.5-GNN model.
- For each time step and node, the procedure applies GNN and GRUcell operations, then uses an MLP to produce predicted outputs.
- The procedure appends each predicted X-hat value to an output list.
- The deployment framework and prediction-curve comparison are presented in Figures 10 and 11.
D NODESFC-GRU
nodesFC-GRU replaces PM2.5-GNN’s GNN with a fully connected MLP to assess neighborhood information, but performs poorly in settings with weak PM2.5 factors because it overfits.
- The proposed PM2.5-GNN consists of a knowledge-enhanced GNN and a spatiotemporal GRU.
- nodesFC-GRU replaces the GNN module with a fully connected MLP to measure the contribution of neighborhood information.The comparison is designed as a degradation experiment within PM2.5-GNN.
- nodesFC-GRU is undesirable when PM2.5 factors have relatively weak effects on the considered location.
- Its large number of parameters in the fully connected module inevitably leads to overfitting.
E CASE STUDY: XI’AN
A fine-grained Xi’an case study compares six models with ground-truth concentrations, and PM2.5-GNN’s prediction curve fits the ground truth best. Across automatic evaluations and case studies, the authors report improved accuracy from effectively leveraging domain knowledge.
- The Xi’an case study evaluates six models against ground-truth PM2.5 concentrations at fine-grained scale.
- PM2.5-GNN’s prediction curve fits the ground-truth curve best in Xi’an.
- Across automatic evaluations and case studies, PM2.5-GNN predicts PM2.5 more accurately by leveraging domain knowledge effectively.