SRM Institute of Science & Technology
Spatiotemporal GNN for PM2.5 Forecasting
Graph neural networks forecasting air quality across 184 cities.
Jan 2022 – May 2022
Framed for
GINs + GRUs beating temporal baselines on a 184-city graph
A spatiotemporal GNN in PyTorch Geometric that models spatial coupling between cities, improving both error and threshold-detection metrics over baseline.
- −7.2%
- RMSE versus baseline
- +12.4%
- CSI versus baseline
- 184
- cities modeled
- PyTorch Geometric
- GNN
- GIN
- GRU
- Python
Choosing a graph model because the problem was actually spatial
Recognized that temporal baselines structurally could not capture inter-city coupling, and selected an architecture matched to the problem's shape, validated on a four-day horizon.
- −7.2%
- RMSE
- 4-day
- prediction horizon
- 184
- cities
- Model Architecture
- Research
- PyTorch
- Graph Modeling
The problem
PM2.5 concentration is spatially coupled (a city's air quality depends on its neighbors'), which purely temporal baseline models cannot represent.
What I built
- 01Engineered a spatiotemporal architecture combining Graph Isomorphism Networks with GRUs in PyTorch Geometric.
- 02Modeled 184 cities as a graph so spatial coupling was learned rather than ignored.
- 03Targeted four-day prediction horizons.
Outcome
7.2% lower RMSE and 12.4% higher CSI than baseline models, at four-day prediction horizons across 184 cities.