Noise robustness

Figures 5 and 6 study how each method degrades under input perturbations: Gaussian noise on the temperature signal and impulse noise that simulates sensor drop-outs.

Run

uv run jupyter lab swissrivernetwork/benchmark/visualize_results/noises.ipynb

The notebook:

1. Loads checkpoints from outputs/ray_results/.
2. Applies Gaussian / impulse noise to the test splits at configurable strengths.
3. Computes RMSE / MAE / NSE degradation curves and writes them to visualize_results/outputs/noises/.
4. Renders the two paper figures.

Custom noise levels

Modify the NOISE_STDS and IMPULSE_PROBS constants at the top of the notebook. Re-run the affected cells — no retraining is required.

Interpretation

• Graph-aware methods (Graphlet, ST-GNN) are typically more robust to per-station noise because neighbor predictions dampen outliers.
• The Transformer-Embedding variant has the flattest degradation curve under impulse noise, consistent with the paper's discussion.