Systematic review of energy theft practices and autonomous detection through artificial intelligence methods

Electricity theft is one of the most difficult forms of energy loss to identify. Unlike technical losses, which arise naturally from the physical behaviour of transmission and distribution systems, non-technical losses are caused by external actions such as illegal connections, meter tampering, bypass circuits, and deliberate manipulation of consumption measurements. The consequences extend far beyond the unpaid electricity itself. Energy theft reduces the revenues of distribution operators, increases costs for legitimate consumers, can damage infrastructure, and may compromise the reliability and safety of the grid. In some cases, fraudulent connections can contribute to supply interruptions, equipment failures, and dangerous electrical conditions. The challenge is made even more complex by the variety of strategies used to conceal illegal consumption. Different users, grid configurations, voltage levels, and metering technologies can generate very different patterns, making it difficult to design a single detection method that works reliably in every situation.

A comprehensive analysis of electricity theft begins with the physical mechanisms used to alter or avoid measurement. In medium-voltage systems, fraudulent users may manipulate the measuring transformers connected to the meter, reducing the current registered by the metering apparatus. Other techniques rely on external electronic boards equipped with microcontrollers, relays, and radio communication systems. These devices can periodically disable parts of the measurement circuit, lowering the recorded consumption without producing an obviously impossible load profile. In low-voltage networks, theft can occur through unauthorised cables connected directly to the distribution system, parallel circuits that bypass the electricity meter, or modifications made inside the meter itself. Permanent magnets may also be used to interfere with current measurement components, while small holes can be drilled into the device to interrupt specific electrical connections. Moreover, electricity theft is not limited to a single technical method. It may involve visible illegal connections, concealed underground cables, sophisticated electronic modifications, or intermittent manipulation designed to resemble ordinary changes in consumption. This diversity explains why conventional inspections alone are insufficient. The same reduction in measured energy may result from fraud, a faulty meter, changes in user behaviour, seasonal effects, or legitimate variations in the connected load.

The deployment of smart meters has changed the possibilities for detecting anomalous consumption. Earlier metering systems often provided only monthly aggregate values, offering limited information about how electricity is used over time. Newer devices can record measurements every fifteen minutes and communicate data directly to distribution operators. This higher temporal resolution creates detailed load profiles that can reveal unusual variations, persistent reductions in base consumption, unexpected changes between tariff periods, and other patterns that may be associated with fraudulent behaviour. However, greater data availability does not automatically solve the problem. Electricity networks can include millions of consumers, producing vast amounts of heterogeneous information. The data may contain noise, missing observations, seasonal changes, and highly imbalanced examples, because confirmed cases of theft are rare compared with legitimate consumption. The real opportunity therefore lies in combining smart metering infrastructures with artificial intelligence capable of interpreting these complex data streams.

Artificial intelligence can support several interconnected functions within an energy supervision system. It can predict future consumption, monitor the evolution of physical quantities, detect and locate anomalies, classify the suspected problem, and support the actions needed to resolve it. For electricity theft, the process begins with data collected from smart meters, sensors, consumer profiles, and the distribution network. These measurements are cleaned, denoised, normalised, filtered, and associated with contextual information such as weather conditions, grid topology, power quality, and historical consumption. Metering data and contextual information are transformed into alerts and reports, while human experts remain involved in validating the output and deciding whether an on-site inspection is necessary. This is an important distinction. Artificial intelligence is not presented as a complete substitute for field technicians or legal verification. Its role is to reduce the enormous search space, identifying the users or areas where further investigation is most justified.

Different machine learning approaches can be used depending on the available data and the type of theft being investigated. Convolutional neural networks are particularly effective at extracting local patterns and relevant features from large datasets. Recurrent architectures can instead model the temporal evolution of electricity consumption, making them suitable for detecting fraud that develops gradually or occurs intermittently over long periods. Hybrid models combining convolutional and recurrent layers are especially promising because they can analyse both the shape of a consumption pattern and its evolution through time. Autoencoders can support anomaly detection when labelled examples are scarce, while federated and distributed learning may allow multiple grid entities to collaborate without centrally collecting sensitive consumption data. Despite their potential, these methods remain strongly dependent on the quality of the training process. Neural networks require suitable architectures, representative datasets, careful hyperparameter selection, and mechanisms to prevent overfitting. A model trained on a specific type of meter manipulation or consumer population may perform poorly when transferred to a different grid.

The central difficulty is not building an algorithm that detects one known theft pattern. It is creating a system that remains reliable across different users, infrastructures, measurement technologies, and fraudulent techniques. Industrial facilities, households, commercial buildings, and prosumers have fundamentally different consumption profiles. The physical topology of the grid also varies, as do communication systems and the quantity of information available to the operator. Even the same theft method can generate different effects depending on the connected load and the way it is applied. Data imbalance represents another major limitation. Confirmed theft cases are relatively rare, while normal consumption data are abundant. Synthetic data generation, resampling, transfer learning, and ensemble models can partially mitigate this issue, but each introduces additional risks related to bias, computational complexity, and poor transfer between different operational contexts. False positives are particularly costly. Incorrectly flagging an honest customer can trigger field inspections, administrative procedures, and legal consequences. For this reason, practical systems must prioritise not only detection accuracy, but also reliability, explainability, and controlled interaction with human experts.

A practical detection strategy can be organised into a sequence of connected stages. The distribution network is first analysed to identify feeders or areas with abnormal non-technical losses. Individual consumption profiles within those areas are then processed by machine learning models to identify suspicious behaviour. Field technicians subsequently inspect the selected installations, and physical evidence is collected to confirm the fraud. The proposed direction is to automate the first two analytical stages as much as possible. Smart meters and monitoring infrastructures provide the measurements, artificial intelligence detects and ranks suspicious cases, and human intervention is reserved for final validation. This approach can reduce inspection costs while increasing the probability that technicians are sent to locations where fraud is genuinely present.

Electricity theft detection is not only an economic problem. Non-technical losses affect grid planning, infrastructure reliability, public safety, and the broader transition toward sustainable energy systems. Artificial intelligence offers a way to transform fragmented measurements into actionable knowledge, but effective implementation requires more than selecting a high-performing algorithm. It demands reliable metering infrastructure, secure communication, careful preprocessing, representative data, adaptable neural architectures, and continued collaboration between utilities, regulators, engineers, and AI specialists. The future of autonomous theft detection will therefore depend on combining physical knowledge of the grid with data-driven learning. The most effective systems will not treat electricity theft as a purely statistical anomaly, but as a complex interaction between human behaviour, metering technology, network structure, and evolving consumption patterns. In this perspective, artificial intelligence becomes the analytical layer of a broader protection framework: one capable of monitoring the grid continuously, identifying suspicious behaviour early, and directing human expertise where it can have the greatest impact.