Forecasting Air Quality with Linear Autoregressive Models

This repository contains the implementation for a final project of the Applied Machine Learning in Python course at LMU Munich, Summer Semester 2025. It focuses on forecasting air quality (specifically NO2 levels) using linear autoregressive models.
It builds on materials from AppliedML, a fork of the original course repository by @mselezniova.

Requirements

To set up the development environment, first create and activate a Python virtual environment:

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

To install the required packages:

pip install -r requirements.txt

Note: This project also depends on components from a forked version of max-georgiev's repo. Please make sure to clone that repository and use it alongside this one, as it provides essential utility functions and __init__.py files for modular imports.

Data

This project uses the UCI Air Quality Dataset, originally from the UCI ML Repository.

The dataset (AirQualityUCI.csv) is already included under data/raw/, so no manual download is necessary.

---

Training

To train a model for forecasting pollutant levels (e.g., NO₂), first ensure you’ve set the correct configuration in src/utils/config.py, including:

TARGET_POLLUTANT (e.g., 'NO2(GT)')

START_DATE and END_DATE

LAG_DEPTH (e.g., 24)

To train and evaluate the models for pollutant forecasting (e.g., NO₂), open and run the following Jupyter notebook:

src/notebooks/main.ipynb

This notebook performs the full pipeline:

• Data loading and preprocessing
• Lagged feature engineering (e.g., using past 24 values)
• Model training (e.g., gradient descent, ridge regression)
• Evaluation using MSE, RMSE, and MAE
• Visualizations such as:
  • Prediction vs. actual values
  • Residual analysis
  • ACF/PACF plots
  • MAE by hour of day or weekday
  • Model coefficient interpretation

Results

The table below shows example test results from forecasting NO₂ (GT) levels using different models and lag depths.
A lag depth of 24 yielded the best performance. Larger depths (e.g., 48) slightly increased test error, likely due to overfitting.

Lag Depth	Model	MSE	MAE	RMSE
1	Courselib LR (GD 0.001)	149.70	13.39	12.24
1	Scikit-learn Ridge (α = 0.1)	149.71	13.39	12.24
2	Courselib LR (GD 0.001)	137.37	13.01	11.72
2	Scikit-learn Ridge (α = 0.1)	137.37	13.01	11.72
24	Courselib LR (GD 0.001)	137.27	12.99	11.72
24	Scikit-learn Ridge (α = 1.0)	137.42	13.01	11.72
24	Scikit-learn Ridge (α = 0.1)	137.25	12.99	11.72
48	Courselib LR (GD 0.001)	140.03	12.98	11.83
48	Scikit-learn Ridge (α = 0.1)	140.07	12.98	11.84

Overall, linear autoregressive models, especially ridge regression, performed well for short-term air quality prediction.
The best-performing model (scikit-learn Ridge with α = 0.1) slightly outperformed both the course-provided gradient descent model and Ridge with α = 1.0.

Example plots (predicted vs actual, error by hour, coefficient weights, etc.) can be found in notebooks/main.ipynb.
The notebook also contains additional analysis and visual reasoning used to guide model selection.

Contributing

This project is licensed under the Apache License 2.0. See the LICENSE file for details.