Federated Extreme Weather Detection

Codebase for the CS-E4740 Federated Learning '26 project.

Overview

This project formulates an unsupervised anomaly detection task over FMI Open Data as a Graph Total Variation Minimization (GTVMin) problem. The focus is to identify extreme weather events (e.g. storms/frost) using autoencoder reconstruction error, where node collaboration (federated learning) assists stations with scarce local anomaly representation.

Data Assumption

Raw FMI weather observations should be downloaded via your load_data.py into a ./fmi_data/ repository as .csv files. The project relies strictly on four derived variables (Air temperature, Relative humidity, Wind speed, Precipitation amount).

Structure

• config/: Contains default_config.yaml specifying alpha, architectures, & graph params.
• src/:
  • data_interface.py: Handles CSV parsing.
  • preprocessing.py: Forms rolling windows and constructs ground truth anomaly labels for test phase.
  • graph.py: Employs correlation-of-temperature as an adjacency metric (System design choice).
  • models/autoencoder.py: PyTorch Autoencoders (Small & Large variants).
  • fl_training.py: Executable loop applying the fixed-point operator synchronous projection step param.grad.add_(2 * alpha * coupling).
  • local_baseline.py: Local PyTorch training decoupled.
  • evaluation.py: Extract predictive errors and evaluates area metrics (ROC, PR).

Execution

1. Make sure PyTorch and dependencies are installed (pip install torch pandas numpy scikit-learn matplotlib pyyaml).
2. Populate data using load_data.py.
3. Run python run_experiments.py to train Systems A, B, and C computationally.
4. Run python analyze_results.py to compare their relative AUC on predicting storms/extreme environments.

Course Mapping Guarantees

• Unsupervised MSE acts as $L_i$.
• Coupling is represented as an explicit graph momentum step utilizing $\alpha$ hyperparameter.
• Validates 3 Systems: Local (System A), FedSmall (System B), FedLarge (System C)