Electricity theft detection is the work of finding energy that bypasses the meter, or passes through a tampered one, from the traces it leaves in consumption data. As of 2026 it increasingly combines automatic meter reading (AMR) data, transformer measurements and customer history, scored with AI, so field crews visit the most suspicious connections first.
Key takeaways
- Electricity theft detection uses traces in meter, AMR and transformer data to find suspect customers and prioritise site inspections.
- According to the TEDAŞ 2024 Sector Report, Türkiye's combined technical and non-technical distribution loss rate was 11.02% in 2023.
- No single signal is proof; consumption breaks, peer gaps, transformer energy balance and meter event logs are weighed together.
- AI does not deliver a "theft" verdict but a risk score; theft is confirmed by an on-site check and a report under EPDK rules.
- The measure of success is the inspection hit rate: out of every 100 customers visited, how many actually produce a finding.
Who deals with electricity theft, and how?
Electricity theft is not only a distribution company problem. Organised industrial zones (OSBs, from the Turkish acronym) that run their own distribution networks, shopping centres and office towers that resell energy to tenants through sub-meters, large housing estates and industrial campuses all face the same question: why does the energy entering the network not match the energy being billed? Municipalities ask the same question about their water networks; other municipal uses are gathered in AI in local government.
The traditional approach relies on tip-offs, random site checks and spotting obvious drops at reading time. This traditional approach has two weaknesses. Field crews are limited, and most of their visits end at clean connections. And as tampering becomes more skilled, the trace gets fainter: crude bypasses that push consumption to zero give way to methods that under-record only at certain hours or by a certain share.
In Türkiye, the definitions of illegal and irregular use, the inspection report and the back-billing rules are set out in the Electricity Market Consumer Services Regulation issued by the Energy Market Regulatory Authority (EPDK). That is why theft analytics software never delivers a verdict of "theft" on its own; it only tells you which customer to inspect first. The legal process starts with the on-site check and a report drawn up in line with the regulation, and the customer keeps the right to object.
How high are distribution losses in Türkiye, and what do they cost?
Türkiye's electricity distribution loss rate remains well above that of many European countries. The official figure combines technical losses (physical losses such as line and transformer heating) with non-technical losses (theft, meter faults and unmeasured consumption):
According to the 2024 Türkiye Electricity Distribution Sector Report by Turkish Electricity Distribution Corporation (TEDAŞ), Türkiye's combined technical and non-technical electricity distribution loss rate was 11.02% in 2023.
The consumption table in the TEDAŞ report makes the loss rate tangible: in 2023, 23,993 GWh of electricity was lost on the distribution grid, against national net consumption of 289,372 GWh. How much of that is theft varies widely by region, but every non-technical kilowatt-hour was generated, transported and never paid for.
A European comparison shows the room for improvement. Citing the Council of European Energy Regulators (CEER) 3rd Report on Power Losses, the TEDAŞ report gives 2022 distribution loss rates of 2.8% in Finland, 4.6% in the Netherlands and 6.7% in Ireland (Türkiye's 11.02% refers to 2023). The same report shows distribution users rising from 43.6 million in 2018 to 49.7 million in 2023; the larger the customer base, the less a manual inspection model can keep up.
At site level the cost of losses looks different but works the same way. In a shopping centre or industrial zone with sub-metering, unmeasured consumption ends up in service charges or network fees, so honest users pay for it rather than the people taking the energy.
Electricity theft detection methods: the signals in the data
Electricity theft detection does not rely on a single indicator but on signals that corroborate one another; technically, it is an anomaly detection problem. The most useful signals are:
| Signal | What you look at | Enough on its own? |
|---|---|---|
| Consumption break | A customer's consumption dropping permanently after a certain date | No; could be a move or a change of business |
| Peer comparison | Difference from customers with the same tariff, sector and connection capacity | No, but strong for prioritisation |
| Transformer energy balance | The gap between energy measured at the transformer and the sum of its customers | Narrows the area, not the customer |
| Meter event logs | Cover-open, magnetic interference, missing phase and reverse-flow alerts | Strong evidence; needs site confirmation |
| Load profile shape | An hourly profile that does not fit the sector or season (e.g. zero at night, low by day) | No; needs remote reading data |
| Ratio anomalies | Unexplained change in the active-to-reactive consumption ratio | Meaningful for industrial customers |
Most of these theft detection signals need frequent, reliable meter data. A meter read by hand once a month has no load profile and no event log. That is why the foundation for theft analytics is usually automatic meter reading plus transformer-level measurement.
How does AI electricity theft detection work? A seven-step framework
The seven-step framework for AI-assisted theft detection works the same way for a distribution area as for an industrial zone or a shopping centre:
- Establish the energy balance. Confirm which transformer, feeder or main meter supplies which customers. If the network topology is wrong, even the best model points to the wrong place; the approach in our article on GIS for municipalities and networks applies directly here.
- Bring the data into one table. Meter readings, event logs, customer records (tariff, sector, connection capacity), past inspection outcomes and field notes must match on the same key.
- Add a rule-based first filter. Clear-cut cases such as a cover-open event or an active customer with zero consumption should become work orders without waiting for a model.
- Train the anomaly model. Models measure how far each customer departs from its own history and its peer group. We explain the general methods in our anomaly detection guide; the difference here is teaching the model about seasonality and tariff changes.
- Give each customer a risk score. The output is not a list but a score with reasons, such as "consumption well below peers since March, with a cover-open event in the same period".
- Plan inspections by score. Work orders are routed starting from the highest-scoring customers, and the on-site report and photos flow back into the system.
- Feed the results back. Every inspection outcome (theft, irregular use, meter fault, clean) is a label. As labels accumulate the model gets sharper, which is why the data labelling process should be designed from day one.
In theft analytics, the measure of success is not the number of thefts found but the inspection hit rate: out of every 100 customers visited, how many actually produce a finding. Measuring that rate with your current method before the project, then comparing, is the cleanest way to measure the ROI of an AI project.
Risk scoring has a second benefit: transparency. When the reason each customer is on the list is written down, crews know what to look for and management can see where field time goes. Over time those reasons become local knowledge that stays with the organisation when an experienced inspector leaves.
Common mistakes in electricity theft analytics
- Looking only for drops. Theft does not always zero out consumption; it may hide only part of it. Without profile and peer analysis these customers stay invisible.
- Mistaking technical losses for theft. Long, heavily loaded lines have high technical losses; blaming the whole transformer gap on theft sends crews to the wrong place. Electricity load forecasting models help separate expected load and losses.
- Not recording inspection outcomes. If the report stays on paper, the model learns nothing.
- Overlooking personal data. Consumption profiles of individual customers count as personal data under KVKK, Türkiye's data protection law; access rights, retention periods and privacy notices belong in your KVKK compliance steps.
If you want to see the theft detection method applied to water, our article on water loss analysis describes a similar approach that narrows down non-revenue water with district meters.
As more meters are read remotely, the reach of theft analytics grows; we summarise the smart meter replacement timetable for sites above 10 MWh a year (as of 2026) in our article on the automatic meter reading obligation in Turkey.
Theft analytics is one part of a wider change covering inventory, metering and field operations. We bring those layers together in our guide to digital transformation in energy utilities.
How Digital Bridge helps with electricity theft detection
Digital Bridge does not start electricity theft projects by installing off-the-shelf software; we start from your data:
- Discovery and needs analysis. We review your metering infrastructure, reading frequency, customer data and network topology. Through our energy and utilities solutions we set out in writing which data is missing and which signals you can use today.
- Measurement infrastructure. If reading frequency or event logging falls short, we install remote reading with our OSOS meter reading system. Because the same team builds the device, the embedded software and the panel, we manage the data flow end to end. Our LoRa vs NB-IoT vs cellular comparison helps with the communications choice.
- Pilot. We start with one transformer area, one section of an industrial zone or one shopping centre. Through our anomaly detection service we produce risk scores, build the first inspection list with your field team and compare the hit rate with your current method.
- Integration. We connect risk scores to your billing system, work order software and maps. Suspect customers appear by transformer on a geographic information system (GIS), and work orders follow the flow described in our field service management guide.
- Reporting. Loss and hit-rate indicators by area, transformer and customer group sit on a BI dashboard; our guide to BI dashboard KPIs explains which indicators matter.
For industrial sites this work often goes hand in hand with efficiency: while you monitor your own meters you can also follow factory energy monitoring and reactive power penalty risk from the same data.
Where should an AI project start?
Theft analytics is one of the AI use cases whose results are easiest to measure: every inspection either finds something or it does not. Even so, the first step is fixing the data, not building the model. Our article on where to start with AI in business covers the general starting logic, and you will find related guides on our Artificial Intelligence topic page.
Next step
To prepare for an electricity theft pilot, put the last 12 months of meter readings and inspection outcomes into a single file; that is the starting point of a pilot. Then contact us: we will review your data together and set out in a written proposal which signals your theft analytics can start with and how to choose the pilot area.