Consumption anomaly screening
Unsupervised screening of smart meter consumption for premises whose pattern has changed or differs from similar customers, used where inspection history is thin to find candidates for inspection, many of which turn out to be faulty meters rather than theft.
Discovery, not a referral queueTechnique: Anomaly Detection
The screening compares each premises’ consumption with its own history and with similar customers, and flags those whose pattern has shifted or sits far from its group.
A revenue protection analyst reviews the flagged premises, discards those explained by vacancy, new generation, or a change of occupant, and sends the rest for inspection.
One widely cited study combined transformer-level metering with detectors of abnormal consumption patterns, using the transformer to shortlist an area before examining individual customers, which keeps the method from being run blind across the whole customer base.
The inspections it triggers become the labelled examples that a supervised model later needs.
This is a discovery tool for the period after a smart meter rollout, not a way to decide who is stealing.
It is the wrong choice where a substantial inspection history exists, since a model trained on inspection outcomes, like the production system Naturgy’s team reports beating its earlier rules, is likely to rank premises more precisely.
Strengths for this problem
- Works without inspection labels, so it is usable soon after smart meters are installed.
- Finds patterns nobody has written a rule for, such as a gradual decline, a change in daily shape, or consumption far below similar premises.
- Surfaces faulty and wrongly configured meters as well as tampering, which are worth correcting in their own right.
Limitations for this problem
- Unusual is not the same as irregular, so a vacant property, a new solar installation, or a change of occupant all look like theft.
- Precision is low, and every candidate needs a person to look at it before an inspector is sent.
- Smart meter data at the granularity the method needs may be restricted for domestic customers by privacy rules.
Data needed
- Interval or daily consumption per premises over at least a year, to capture seasonal patterns.
- Customer class, tariff, premises type, and any known solar or other generation, to define comparable groups.
- Change-of-occupancy, vacancy, and meter exchange records, to explain legitimate changes.
How it is measured
- Share of screened candidates that an analyst sends for inspection, and the inspection hit rate of those.
- Split of findings into theft, faulty meters, and configuration or billing errors.
- Number of candidates that become labelled examples for an inspection targeting model.
Pitfalls
- Sending every anomaly to the field, which spends inspection capacity on vacant properties and new solar panels.
- Comparing a premises with the whole customer base rather than with customers of the same class and premises type.
- Using detailed domestic consumption data without the consent or legal basis the market's privacy rules require.
Human oversight
A revenue protection analyst reviews each candidate against occupancy, generation, and meter records before deciding whether to send an inspector. Inspectors confirm every finding on site. The screening never produces a bill, a disconnection, or an accusation on its own.