Heatwave Prediction in Switzerland (2025)

Context

This project explores the prediction of Level 2 heatwave episodes in Switzerland, motivated by the rise in deadly heatwaves (1,444 deaths in France in 2025). The goal is to anticipate these events to inform at-risk populations and activate preventive measures.

Methodology

• Data: Utilization of meteorological data from the Pully station (MeteoSwiss) and the NAO (North Atlantic Oscillation).
• Feature Engineering: Creation of derived variables (moving averages, humidex, linear trends).
• Models: Comparison of two approaches:
  • XGBoost: A decision tree-based model.
  • LSTM: A recurrent neural network to capture data temporality.

Results

• XGBoost: F1 score of 0.51 (better than random, but overly alarmist).
• LSTM: F1 score of 0.41 (suffers from overfitting due to limited data).

Conclusion and Perspectives

The results highlight the importance of feature engineering and sufficient data for complex models. Improvements are possible by integrating more stations and geopotential levels.

📖 For more details, read the full article on Medium:
[Link to Medium article](https://www.twine.net/signin