---
license: mit
task_categories:
  - time-series-forecasting
language: []
tags:
  - air-quality
  - spatio-temporal
  - graph-neural-networks
  - missing-data
pretty_name: AirQualityBench
size_categories:
  - 1G-10G
---

# AirQualityBench

A global-scale air quality forecasting benchmark featuring **3,720 monitoring stations** across the world with **authentic missing patterns** and physical-scale evaluation.

## Dataset Summary

AirQualityBench provides a realistic testbed for air quality prediction by preserving raw observational characteristics that prior benchmarks have removed through imputation. The dataset spans **5 years (2021–2025)** of hourly measurements across **6 primary pollutants** (PM2.5, PM10, NO₂, O₃, SO₂, CO).

### Key Features

- **Global scale**: 3,720 active stations spanning 7 continents, sourced from [OpenAQ](https://openaq.org/)
- **Authentic missingness**: No imputation — evaluation uses boolean masks to score only valid observations
- **Physical-scale metrics**: All statistics computed in original concentration units (μg/m³ for gases, mg/m³ for CO)
- **Multi-pollutant**: 6 pollutants evaluated jointly

## Data Structure

Each yearly HDF5 file (`aq_compact_{year}.h5`) contains:

| Dataset       | Shape          | Description                          |
|--------------|----------------|--------------------------------------|
| `values`     | `(T, 3720, 6)`  | Hourly concentrations (float32)      |
| `masks`      | `(T, 3720, 6)`  | Valid observation flags (bool)       |
| `station_ids`| `(3720,)`       | Station identifiers                  |
| `params`     | `(6,)`          | Pollutant names                      |

Additional files:
- `adj_mx_10.pkl` — KNN adjacency matrix (k=10, Haversine distance)
- `scaler.csv` — Per-pollutant mean/std for normalization
- `global_quality_stats.csv` — Per-station, per-pollutant coverage rates
- `selected_nodes_metadata.csv` — Station coordinates and metadata
- `all_stations_coords.csv` — Raw station coordinates from OpenAQ

## Data Splits

| Split      | Years       | Approx. Hours |
|-----------|-------------|---------------|
| Train     | 2021–2023   | ~26,280       |
| Validation| 2024        | ~8,784        |
| Test      | 2025        | ~8,760        |

## Usage

This dataset uses the HDF5 format with multiple internal arrays of different shapes, which is incompatible with the `datasets` library's HDF5 loader. Use `h5py` to read the files directly:

```python
import h5py

with h5py.File("aq_compact_2021.h5", "r") as f:
    values = f["values"][:]       # (T, 3720, 6) float32
    masks = f["masks"][:]         # (T, 3720, 6) bool
    params = f["params"][:]       # (6,) pollutant names
    station_ids = f["station_ids"][:]  # (3720,) station IDs
```

## Benchmark Code

Training and evaluation code: [github.com/Star-Learning/AirQualityBench](https://github.com/Star-Learning/AirQualityBench)

## Citation

```bibtex
@article{xu2026airqualitybench,
  title={AirQualityBench: A Realistic Evaluation Benchmark for Global Air Quality Forecasting},
  author={Xu, Xing and Wang, Xu and Zhang, Yudong and Zhao, Huilin and Zhou, Zhengyang and Wang, Yang},
  journal={arXiv preprint arXiv:2605.05854},
  year={2026}
}
```