In depthAbout a 5-minute read
Meters disagree with inverters, alarms trace to nothing: pass data governance before PV AI
In one real cross-check, daily, monthly, and yearly generation all looked fine — while lifetime totals diverged by -10.72%. Ungoverned data means no model can produce an answer that survives acceptance.
01
When AI answers wrong, the data is usually lying
Many teams start by asking an AI assistant the same question: was generation normal this week? The answer reads fluently, but the numbers disagree with the monthly report, and the team concludes the AI is not ready. Dig deeper and the problem is upstream: three different definitions of the same generation metric across meters, inverters, and settlement; history gaps; missing fields nobody annotated.
Feed contradictory inputs to any model and the output cannot be verified. A fluent answer that fails acceptance kills the pilot in week one — and the failure gets blamed on AI, not on the data.
02
A real cross-check: short-term normal, lifetime off by 10.72%
We ran a meter-versus-inverter cross-check on one plant (sanitized): four windows — day, month, year, lifetime — against a 2% deviation threshold. The three short windows all passed. Lifetime totals diverged by -10.72%, and the system recommended inspecting the metering circuit or auditing line losses, rather than reporting a vague anomaly.
The value is not the number itself but the fact that it ships with its formula, threshold, and reasoning: how the deviation was computed, which window breached, what to check next. That is what governance looks like — not scrubbing data, but making every number explainable.
03
The governance checklist is three lines long
Unified definitions: how each metric is computed and which source wins, in writing. Explicit boundaries: which plants, devices, and time windows are usable, with missing data annotated with verified reasons. Named owners: who confirms each class of data and who arbitrates disputes.
None of this requires replacing systems or a six-month program. Monitoring, reports, and tickets all stay — governance adds a layer of ground truth on top of them.
04
The ten-minute test
Pick one alarm from yesterday at random. Trace it to the device, then to the power curve and handling record at that moment. If the walk takes under ten minutes, the foundation is ready for a first pilot. If it stalls, the step where it stalls is exactly where a data health check should begin.
The test has a second use: when management asks why AI needs a data health check first, run it live. It beats any briefing deck.
05
Governance and AI are the same project
ZenovaOS AI builds governance in as the first layer of AI operations: connect the existing monitoring account (nothing replaced), run the data health check — completeness, definitions, quality baseline — then enable scenarios one by one. The -10.72% cross-check above is a standard output of that health-check package: 4-8 weeks, ending in a prioritized data-quality inventory.
You do not need to finish governing data before considering AI. The health check is the first AI scenario.
Governance only needs to be sufficient for the first pilot scenario — and the ten-minute trace test tells you whether it is.