Energy utility: 50,000 sensors, real-time detection, €2.4M saved.

Eight hours of delay to detect an anomaly on an electrical grid is eight hours during which you pay for the energy leaking.

The context

French energy company, 4,800 employees, Paris hub. Medium-voltage distribution network instrumented with 50,000 sensors (consumption, voltage, harmonics, transformer temperature). Overnight batch data pipeline on Hadoop, anomaly detection via static rules.

The problem

• Anomaly detection with 8h average delay (H+24 batch)
• 30% of incidents detected after client impact
• Unidentified grid losses estimated at 3.8M EUR/year
• Static rules generating 70% false positives
• Limited bandwidth on some industrial gateways (4G, sometimes 2G)

The approach

Edge ML pipeline: compact ONNX models deployed on industrial gateways for local inference in less than 5 seconds. Cloud fallback only for ambiguous cases. Continuous drift detection and automatic monthly retraining.

The architecture

• Isolation forest + autoencoder models quantized ONNX (8 MB)
• Inference on Edge TPU embedded in gateways
• Kafka streaming to Flink for regional aggregations
• MLflow for model versioning, auto retraining if drift > threshold
• Cloud fallback (5% of traffic) for ambiguous cases, decision under 800 ms

The stack

• ONNX Runtime, INT8 quantized models
• Coral Edge TPU on industrial gateways
• Apache Kafka 3.7, Flink 1.18 for streaming
• MLflow for ML lifecycle, Evidently AI for drift detection
• Weekly retraining pipeline on AWS SageMaker

The results

1. Anomaly detection: 8h to less than 5 seconds
2. Undetected incidents: -70%
3. False positives: 70% to 11%
4. Grid loss savings: 2.4M EUR/year validated at 12 months
5. Bandwidth consumed: -82% (local inference)

What we learned

Drift detection is as critical as inference: without it, the model degrades silently. INT8 quantization loses 1.8 precision points, acceptable here but to validate case by case. Mistake: we wanted to deploy to all 50,000 sensors in six months, we had to stretch to 10. To redo: pilot on 500 sensors for 8 weeks before any rollout, that's what saved the project.