Energy and utilities
Load forecasting, grid operation and maintenance: where AI holds in utilities and why critical infrastructure carries its own set of duties.
What this is about
| Task | Classification |
|---|---|
| Day-ahead load forecast | unproblematic |
| Generation forecast from weather data | unproblematic |
| Condition monitoring of plant | unproblematic |
| Predictive maintenance planning | unproblematic |
| Triaging fault reports | limited |
| Controlling protection or load shedding | high risk, critical infrastructure |
| Per-household tariff recommendations from consumption profiles | limited to high, personal data |
Where it pays off fastest
- 01
Load forecasting
Historical load, calendar, weather. One percentage point less forecast error feeds straight through to procurement cost.
- 02
Generation forecasting
For wind and solar. Quality depends on the weather source, not on the model.
- 03
Condition monitoring
Vibration, temperature, partial discharge. A failure spotted two days earlier is a planned outage instead of an unplanned one.
- 04
Fault intake
Classify reports, propose urgency, prepare dispatch. Always with a check by the control room.
The mistake that almost always happens
A model is trained on weather data known in hindsight but not available at prediction time. It looks excellent in testing and collapses in operation. See Data leakage.
- Use only data that genuinely existed at prediction time, checked by timestamp.
- Validate by time, not at random. A random split over a time series is worthless.
- Measure against the simplest rule: yesterday at the same hour. Beat that or you have gained nothing.
- Plan for seasonal shifts. A model from summer does not describe winter.
Critical infrastructure under the AI Act
Annex III lists systems used as safety components in the operation and management of critical infrastructure, expressly including the supply of electricity, gas and heat. What decides it is the function as a safety component, not the industry.
| Use | Safety component? | Consequence |
|---|---|---|
| Forecast as a planning basis | no | General duties |
| Switching proposal, person decides | as a rule no | Document the oversight |
| Automatic load-shedding control | yes | Full high-risk duties |
| Protection function in plant control | yes | Full high-risk duties |
On top of that come the duties from NIS2 and, for operators of critical installations, from national law. Both regimes run in parallel and carry their own reporting deadlines. See Reporting problems.
Meter data
A quarter-hour reading is a behavioural record. It shows when someone gets up, cooks, showers, works and travels.
- Examine the legal basis separately per analysis purpose. Billing does not carry profiling.
- Aggregate before analysing wherever the purpose allows. A grid state needs no household resolution.
- Set retention by the billing obligation, not by what is technically possible.
- For automated tariff or disconnection decisions, examine Art. 22 GDPR.
Operation
- Measure forecast quality continuously against the naive rule, not only against last year's model.
- Watch for distribution shift: new generation, new tariffs and changed consumption devalue a model faster than age does.
- Keep a fallback procedure and rehearse it. A forecasting system that fails must not block trading.
- Log the model version and data state per decision, see Logging.
Entry sheet: energy and utilities
FREE ACCOUNT
Entry sheet for energy and utilities
This sector's tasks by risk class, the points to settle beforehand, and the metrics the benefit shows up in.
Checklist3 items