ARotting/chronos-microgru

Chronos MicroGRU

Chronos MicroGRU predicts the next eight industrial-telemetry timestamps from a 32-step context. A compact recurrent network produces both means and Gaussian variances for six channels.

Training uses only anomaly-free windows from devices 0-7. Variance scaling is selected on devices 8-9, and final metrics come from unseen devices 10-11. A persistence forecast that repeats the final observed value provides the control.

Reproduce

uv run python projects/edge-sentinel-ml/generate_data.py
uv run python projects/chronos-microgru/train.py

Verified results

• Parameters: 8,208
• Context: 32 timestamps
• Forecast horizon: 8 timestamps
• Normal training windows: 3,830
• Held-out test windows from devices 10-11: 961
• Model normalized RMSE: 0.2663
• Persistence normalized RMSE: 0.3730
• RMSE improvement: 28.61%
• Nominal 90% interval coverage: 88.63%
• Validation-selected variance scale: 1.05

The recurrent model beat persistence in RMSE and MAE for all six channels. Original-unit RMSEs were 1.008 temperature units, 0.820 pressure units, 0.087 vibration units, 0.301 current units, 0.899 flow units, and 4.042 packet-rate units.