Air Quality Prediction - Poprad, Slovakia

An air quality prediction system designed to forecast AQI (Air Quality Index) for Poprad, Slovakia using historical air‑quality and meteorological data. The system ingests historical air-pollution and weather data, trains ML models (e.g. XGBoost and LSTM), and serves a predictive model (with a live demo on Hugging Face Spaces). The workflow is fully automated using GitHub Actions: data collection, model training, and deployment all run as scheduled or on-demand workflows.

🌍 Demo

🔗 Live Demo: Poprad AQI Prediction on Hugging Face

📋 Table of Contents

• Key Features
• Methodology
• Results and Insights
• Automation Pipeline
• Getting Started

🌟 Key Features

• Time Series Recursive Forecasting: Utilizes an XGBoost model for accurate forecasting.
• Containerized: Fully containerized with Docker for easy setup and consistent environments.
• Automated Pipeline: CI/CD pipeline with GitHub Actions for automated testing and building.
• Data-Driven: Uses historical data to train the prediction model.

🔬 Methodology

About Poprad

Poprad is located in northern Slovakia, near the High Tatras mountains. The city's air quality is influenced by:

• Geographic Location: Valley position affects air circulation
• Industrial Activity: Local manufacturing and transportation
• Seasonal Variations: Winter heating and summer tourism patterns
• Weather Patterns: Mountain weather systems and temperature inversions

Data Collection

The project utilizes various environmental parameters that influence air quality:

• Particulate Matter: PM2.5, PM10 concentrations
• Gaseous Pollutants: NO2, SO2, CO levels
• Meteorological Data: Temperature, humidity, wind speed, atmospheric pressure
• Temporal Features: Hour, day, month, season

Data Sources

The project integrates data from multiple sources:

• Slovak Hydrometeorological Institute (SHMÚ): Official air quality monitoring stations
• European Environment Agency: Regional air quality data
• OpenWeatherMap API: Meteorological parameters
• Local monitoring stations: Real-time sensor data

Machine Learning Models

The project implements and compares several ML algorithms:

• Gradient Boosting: Advanced ensemble technique (XGBoost/LightGBM)
• Long Short-Term Memory (LSTM): For time series prediction

Inference

• Prediction Window of 3 days - based on last 3 days (inputs), model predicts next 3 days (targets)
• Recursive Forecasting - a multi-step time series forecasting method where a model trained for one-step-ahead prediction is used iteratively to generate forecasts for multiple steps into the future

Evaluation

• Using single metric for model comparison - The Willmott index - it gives credit for correlation but heavily penalizes systematic errors that would make the forecasts unreliable for air quality management.
• With recursive forecasting error is cumulative - cannot improve Day 3 prediction without improving Day 1 - so it is enough to evaluate only last day's results for simplicity.
• Instead evaluate each output separately (PM2.5, PM10, NO2, SO2, CO) as they tend to differ significantly.

📈 Results and Insights

Model Performance

The Willmott index (Index of Agreement) based on last day's predictions. Higher the better.

Model	PM2.5	PM10	NO2	SO2	CO
XGBoost	0.3716	0.3599	0.4726	0.7681	0.211
LSTM	0.2846	0.4156	0.4476	0.6584	0.2417

Key Findings

• Seasonal Patterns: Higher AQI values during winter months due to heating
• Weather Dependencies: Strong correlation with wind speed and atmospheric pressure
• Daily Cycles: Peak pollution during morning and evening rush hours
• Feature Importance: PM2.5 and meteorological conditions are primary predictors

🤖 Automation Pipeline

The core pipeline consists of the following automated steps, orchestrated by GitHub Actions:

• Docker Image Build: On every code push to main, the Build and Push Docker Image workflow builds a Docker container containing the project environment and pushes it to GitHub Container Registry. This ensures a consistent runtime for all subsequent steps.

• Data Fetch (Hourly): A Fetch Data Hourly workflow runs on a cron schedule (hourly) to update the dataset. It pulls the latest Docker image and executes scripts/fetch_data.py inside the container, collecting new AQI and meteorological data via the APIs. This automated job keeps the dataset current without manual intervention.

• Model Training (Weekly): A Train Model Weekly workflow runs on a weekly cron schedule (e.g. every Monday at 00:00 UTC). It pulls the Docker image and runs scripts/train_model.py in the container to preprocess data, train the models, and evaluate performance. By automating training, the model stays up-to-date with incoming data.

• Model Deployment: The Deploy Model workflow can be triggered manually. It logs into the registry, pulls the latest image, and runs scripts/deploy_model.py inside the container. This step makes trained model available for predictions and updates the live application.

Each workflow reuses the same Docker image to ensure consistency. In summary, code changes trigger a new container build, a scheduled job fetches fresh data hourly, the model is retrained weekly, and an on-demand job deploys the updated model.

🚀 Getting Started

Project Structure

Air-Quality-Prediction/
├── .github/workflows      # GitHub Workflows for automation
├── data/                  # Dataset files
├── notebooks/             # Jupyter notebooks for EDA and modeling
├── scripts/               # Entry points for Docker, GitHub and Hopsworks
├── src/                   # Source code
│   ├── data               # Data loading and preprocessing 
│   ├── hopsworks          # Hopsworks' Feature Store, Model Registry and Deployment
│   ├── model              # Model training and evaluation
├── .dockerignore          # Specify code that should not be copied to Docker Image
├── Dockerfile             # Definition of Docker image used for automation
├── requirements.txt       # Python dependencies

Prerequisites

• Python 3.8+
• pip package manager

Installation

1. Clone the repository:

git clone https://github.com/MartinRajniak/Air-Quality-Prediction.git
cd Air-Quality-Prediction

2. Create a virtual environment:

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies:

pip install -r requirements.txt

Environment Variables

Create a .env file with the following variables:

AQI_TOKEN=your_aqi_api_key
HOPSWORKS_AQI_TOKEN=your_hopsworks_api_key

Usage

Build Docker Image

./scripts/docker_build.sh

Run Any Python Automation Script

./scripts/docker_run.sh scripts/***.py

Contributing

Contributions are welcome! Please follow these steps:

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Commit your changes (git commit -m 'Add amazing feature')
4. Push to the branch (git push origin feature/amazing-feature)
5. Open a Pull Request

📝 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• Slovak Hydrometeorological Institute for providing air quality data
• World Air Quality Index Project for providing the API
• European Environment Agency for regional environmental data
• Open-source community for machine learning libraries
• Contributors and collaborators

---

For questions, issues, or suggestions, please open an issue on GitHub or contact the maintainer directly.