cosuleabianca/eea-pm25-forecasting

EEA PM2.5 Air Quality Dataset

Hourly air quality measurements from the European Environment Agency (EEA) for PM2.5 forecasting research.

Dataset Description

This dataset contains hourly air pollutant concentrations and meteorological data from monitoring stations across 5 European cities, prepared for machine learning forecasting tasks.

Data Sources

• Air Quality: European Environment Agency (EEA) Air Quality Portal
• Weather: Open-Meteo Archive API

Coverage

• Time Period: 2018-01-08 to 2024-12-31
• Countries: 5 (AT, BE, ES, FI, FR)
• Cities: Wien, Paris, Madrid, Antwerpen, Helsinki
• Monitoring Stations: 38
• Total Records: 1,945,153 hourly observations
• Total Features: 81 columns

Dataset Files

Raw Data (Parquet)

File	Description
PM2.5.parquet	PM2.5 concentrations (all sites)
PM2.5_filtered.parquet	PM2.5 (filtered to quality sites)
NO2.parquet	NO2 concentrations (all sites)
NO2_filtered.parquet	NO2 (filtered sites)
PM10.parquet	PM10 concentrations (all sites)
PM10_filtered.parquet	PM10 (filtered sites)

ML-Ready Dataset

File	Description	Size
ml_ready_dataset_full_realistic.csv	Feature-engineered dataset	~1.6 GB

Features (81 columns)

Metadata

• Start: Original timestamp
• Country: Country code (AT, BE, ES, FI, FR)
• SiteNumber: Station identifier
• dt_utc: Timestamp in UTC
• dt_local: Timestamp in local timezone

Site Metadata (Geographic)

• Latitude, Longitude, Altitude

Station Type (One-Hot Encoded)

• StationType_background
• StationType_industrial
• StationType_traffic

Station Area (One-Hot Encoded)

• StationArea_rural
• StationArea_rural-nearcity
• StationArea_suburban
• StationArea_urban

Weather Features (Open-Meteo API)

• temperature_2m: Air temperature at 2m (°C)
• relative_humidity_2m: Relative humidity (%)
• dew_point_2m: Dew point temperature (°C)
• wind_u: East-west wind component (m/s)
• wind_v: North-south wind component (m/s)
• precipitation: Hourly precipitation (mm)
• surface_pressure: Surface pressure (hPa)

Target Variable

• PM2.5: Current PM2.5 concentration (µg/m³)

Pollutant Features

• NO2: Current NO2 concentration
• PM10: Current PM10 concentration

Temporal Features

• hour, day_of_week, day_of_month, month, year
• is_weekend: Weekend indicator (0/1)
• season: Season indicator
• hour_sin, hour_cos: Cyclical hour encoding
• month_sin, month_cos: Cyclical month encoding

Lag Features (1h, 2h, 3h, 6h, 12h, 24h, 168h)

• PM2.5_lag_1h, PM2.5_lag_2h, PM2.5_lag_3h, PM2.5_lag_6h, PM2.5_lag_12h, PM2.5_lag_24h, PM2.5_lag_168h
• NO2_lag_1h, NO2_lag_2h, NO2_lag_3h, NO2_lag_6h, NO2_lag_12h, NO2_lag_24h, NO2_lag_168h
• PM10_lag_1h, PM10_lag_2h, PM10_lag_3h, PM10_lag_6h, PM10_lag_12h, PM10_lag_24h, PM10_lag_168h

Rolling Mean Features (3h, 6h, 12h, 24h windows)

• PM2.5_rolling_mean_3h, PM2.5_rolling_mean_6h, PM2.5_rolling_mean_12h, PM2.5_rolling_mean_24h
• NO2_rolling_mean_3h, NO2_rolling_mean_6h, NO2_rolling_mean_12h, NO2_rolling_mean_24h
• PM10_rolling_mean_3h, PM10_rolling_mean_6h, PM10_rolling_mean_12h, PM10_rolling_mean_24h

Rolling Std Features (3h, 6h, 12h, 24h windows)

• PM2.5_rolling_std_3h, PM2.5_rolling_std_6h, PM2.5_rolling_std_12h, PM2.5_rolling_std_24h
• NO2_rolling_std_3h, NO2_rolling_std_6h, NO2_rolling_std_12h, NO2_rolling_std_24h
• PM10_rolling_std_3h, PM10_rolling_std_6h, PM10_rolling_std_12h, PM10_rolling_std_24h

Data Quality

Filtering Criteria

Stations included meet these quality thresholds:

• Train completeness: ≥50% (2018-2022)
• Test completeness: ≥50% (2023-2024)
• Maximum gap: ≤168 hours

Preprocessing

• Sentinel values (<0) replaced with NaN
• Time-based lag/rolling features (handles data gaps correctly)
• Weather data merged by nearest hour
• Local timezone conversion for temporal features
• No missing values in final dataset

Stations by Country

Country	City	Stations
AT	Wien	10
BE	Antwerpen	8
ES	Madrid	9
FI	Helsinki	5
FR	Paris	6

Usage

Load with Pandas

import pandas as pd

# Load ML-ready dataset
df = pd.read_csv("ml_ready_dataset_full_realistic.csv")

# Train/test split (temporal)
train = df[df['dt_utc'] < '2023-01-01']
test = df[df['dt_utc'] >= '2023-01-01']

Load with Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("cosuleabianca/eea-pm25-dataset")

Load Raw Parquet Files

import pandas as pd

pm25 = pd.read_parquet("PM2.5_filtered.parquet")
no2 = pd.read_parquet("NO2_filtered.parquet")

Train/Test Split

Split	Period	Purpose
Train	2018-01-08 to 2022-12-31	Model training
Test	2023-01-01 to 2024-12-31	Evaluation

This temporal split simulates real-world forecasting scenarios.

Regenerating the Dataset

If you prefer to regenerate from raw EEA data:

# Clone the repository
git clone https://github.com/CosuleaBianca/eea-pm25
cd eea-pm25

# Install dependencies
pip install -r requirements.txt

# Run data pipeline
python dataset_build/src/download_pollutants.py
python dataset_build/src/filter_pm25_sites.py
python dataset_build/src/process_data.py
python dataset_build/src/prepare_ml_dataset.py
python dataset_build/src/coverage_only_v6.py
python dataset_build/src/dataset_full_realistic_v6.py

Citation

If you use this dataset, please cite:

@misc{eea-pm25-dataset,
  author = {Chisilev Bianca-Iuliana},
  title = {EEA PM2.5 Air Quality Dataset for Europe},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/cosuleabianca/eea-pm25-dataset}
}

Links

• GitHub Repository: Github repository
• Pre-trained Models: Models

License

CC BY 4.0 - You are free to share and adapt, with attribution.

Original data from the European Environment Agency is provided under the EEA standard reuse policy.