Recovering $3.1M in lost generation across 43 renewable sites

The foundation was ingestion: MQTT and API connectors normalizing telemetry from every inverter, turbine, and meter vendor into a single time-series store with one canonical data model. Where vendors offered no API — two of the seven — we parsed their scheduled email exports rather than let sites stay dark.

On top of clean data, the anomaly engine compares each asset against its own weather-adjusted expected output, not naive nameplate capacity. A solar string producing 91% of expectation on a hazy day is fine; the same reading under clear sky pages someone. Detection logic was tuned side-by-side with the operations team to keep alert precision above 85% — because an alarm system operators mute is decoration.

The final layer was workflow: anomalies open tickets with expected-versus-actual revenue impact attached, so the maintenance queue self-prioritizes by dollars, not by which portal someone happened to check that morning.