Source-linked AI summary
The Challenge of Non-Technical Loss Detection using Artificial Intelligence: A Survey
Patrick Glauner, Jorge Augusto Meira, Petko Valtchev, Radu State, Franck Bettinger
TL;DR
NTL detection matters because theft, faulty meters, and billing errors impose substantial economic and grid-related costs, while academic research has recently been sparse. This survey reviews AI-based detection methods, features, and data sets, then identifies challenges and future approaches. It reports broad variation in reviewed data sets and metrics and highlights six major open challenges, including data quality, covariate shift, scalability, imbalance, evaluation, and temporal feature modeling.
Problem
NTLs can reach 40% of distributed electricity, yet few academic NTL detection studies appeared in the last three to four years despite active industrial R&D.
Method
The survey reviews AI-based NTL detection research, including expert systems, machine-learning models, algorithms, features, data sets, and performance measures.
Results
Reviewed data sets range from 30 to 700K customers, while reported accuracy ranges from 0.45 to 0.99 and recall from 0.29 to 1.
Takeaways & Limitations
The survey identifies six open challenges: class imbalance and evaluation metrics, feature description, incorrect inspection results, covariate shift, Big Data scalability, and temporal behavior modeling.
Takeaways & Limitations
Inspection labels can be incorrect, including fraudsters labeled non-fraudulent and non-fraudulent behavior labeled fraudulent.
Abstract
from arXiv · showhide
Detection of non-technical losses (NTL) which include electricity theft, faulty meters or billing errors has attracted increasing attention from researchers in electrical engineering and computer science. NTLs cause significant harm to the economy, as in some countries they may range up to 40% of the total electricity distributed. The predominant research direction is employing artificial intelligence to predict whether a customer causes NTL. This paper first provides an overview of how NTLs are defined and their impact on economies, which include loss of revenue and profit of electricity providers and decrease of the stability and reliability of electrical power grids. It then surveys the state-of-the-art research efforts in a up-to-date and comprehensive review of algorithms, features and data sets used. It finally identifies the key scientific and engineering challenges in NTL detection and suggests how they could be addressed in the future.
1. Introduction
Non-technical losses encompass theft, meter problems, and billing errors that damage providers, grid reliability, resources, and economies. The survey reviews AI-based detection research and identifies unresolved challenges and future directions.
- NTLs include electricity theft, faulty or broken meters, and technical or human errors in meter readings, processing, and billing.
- NTLs can reach 40% of distributed electricity in some countries and cost providers and economies billions of dollars.India’s estimated loss is US$ 4.5 billion, while UK and US estimates range from US$ 1-6 billion.
- NTLs reduce electricity providers’ revenue and profit, grid stability and reliability, and the efficient use of limited natural resources.
- Only few NTL detection studies were reported in the previous three to four years despite active industrial R&D, while academic research declined.
- Energy-balance detection requires network topology and synchronized readings, but changing infrastructure, breakdowns, and measurement timing constrain its use in emerging economies.
- The survey reviews AI-based NTL research, identifies key unsolved challenges, and proposes future methods while connecting them to broader AI problems.
2. The State of the Art
NTL detection research spans broad feature types and methods, but prior surveys offered only narrow comparisons. The reviewed literature uses consumption histories, smart-meter data, customer master data, and payment-related information.
- Prior surveys provide narrow comparisons, whereas this survey reviews a wider range of NTL detection results and derives unsolved challenges.
- Monthly consumption: Traditional-meter studies commonly derive daily average kWh consumption features from monthly or annual readings.For customer m, the feature uses consumption increases between consecutive readings divided by the number of days between them.
- Monthly consumption: Feature windows include previous 24 months, previous six months, and customer or residential-area consumption statistics.Reported statistics include average consumption, maximum consumption, standard deviation, inspection counts, and residential-area averages.
- Smart meter consumption: Smart meters enable short-interval features, with studies using 15-minute or 30-minute consumption measurements.
- Smart meter consumption: Some studies encode 96 four-per-hour measurements into 32 dimensions using weighted sums of three successive binary values.The three values receive weights 4, 2, and 1.
- Master data: Master-data features include location, customer or business type, voltage, phases, meter type, tariff, climate, weather, and contracted power.Credit-worthiness ranking additionally summarizes payment performance, income, and neighborhood prosperity on a 0-to-5 scale.
2.2. Expert systems and fuzzy systems
Expert and fuzzy systems incorporate filtering, clustering, human knowledge, and learned membership parameters for NTL detection. Reported performance varies substantially across datasets and evaluation metrics.
- Expert systems and fuzzy systems: An ensemble pre-filters customers into regular and irregular groups to reduce noise in inspection labels before neuro-fuzzy classification.The system obtains test precision 0.512 and accuracy 0.682.
- Expert systems and fuzzy systems: A fuzzy-logic system using 30-minute readings detects electricity theft and abnormalities among Malaysian low- and high-voltage customers.The reported test recall is 0.55, and the system incorporates human expert knowledge.
- Expert systems and fuzzy systems: Combining an SVM with fuzzy-logic postprocessing integrates human expert knowledge into decisions for approximately 100K customers.The reported test recall is 0.72.
- Expert systems and fuzzy systems: Fuzzy c-means clustering uses six-month consumption features and customer-area information to classify customers through fuzzy membership values.The reported average test precision, called average assertiveness, is 0.745.
2.3. Neural networks
Neural-network approaches range from extreme learning machines and ensembles to self-organizing maps and supervised models using consumption and customer attributes. Their reported results differ across datasets, metrics, and inspection rules.
- Neural networks: Extreme learning machines apply one-hidden-layer neural networks to 30-minute meter readings and report test accuracy 0.5461.
- Neural networks: An ensemble of five neural networks trained on approximately 20K customers combines consumption features with customer-specific attributes.The test precision is 0.626 and accuracy is 0.686.
- Neural networks: Self-organizing maps cluster weekly 15-minute customer data into fraud or non-fraud behavior groups.Inspections occur only when hand-crafted criteria, including cluster fit and absence of contractual changes, are satisfied.
- Neural networks: 0.9267 test accuracy, 0.8526 test precision, and 0.9779 test recall are reported for the self-organizing-map approach.
- Neural networks: A neural network using 12-month average consumption and customer features reports test accuracy 0.8717, precision 0.6503, and recall 0.2947.
2.4. Support vector machines
Support vector machines are applied to highly imbalanced customer datasets using monthly, daily, and industrial consumption features. Reported performance varies with data representation, dataset, and competing algorithm.
- Support vector machines: 0.984 test accuracy is reported for SVM classification of 1,350 Indian customer profiles with 15-minute meter readings.A later 96-to-32-dimensional encoding produces test accuracy 0.92 while simplifying and accelerating training.
- Support vector machines: 0.9628 test accuracy is reported for SVMs on 5K Brazilian industrial customer profiles, versus 0.9620 for KNN and 0.9448 for a neural network.
- Support vector machines: A high-performance-computing extension of the encoded SVM algorithms reports test accuracy 0.89.
2.5. Genetic algorithms
Genetic algorithms are used in NTL detection to optimize SVM hyperparameters and to address inspection-sample imbalance through stratified sampling.
- A genetic SVM globally optimizes the SVM’s Lagrangian multipliers, regularization factor C, and Gaussian kernel parameter γ.
- The main challenge in a 1.1M-customer data set is the much smaller class of inspected customers.
- Stratified sampling increases the number of inspections and minimizes their statistical variance through a nonlinear restricted optimization problem.
- The optimization minimizes overall energy loss due to electricity theft using genetic algorithms and simulated annealing.
- The genetic-algorithm approach outperforms simulated annealing, but only reduced variance is reported and is not comparable with other studies.
2.6. Rough sets
Rough set analysis approximates concepts distinguishing fraudulent from regular electricity use, but reported applications vary substantially in detail and performance.
- Rough sets provide lower and upper approximations of an original conventional or crisp set.
- An application to 40K customers lacks details about the attributes used per customer and achieves a test accuracy of 0.2.
- Supervised rough set analysis approximates concepts describing fraud and regular electricity use.
2.7. Other methods
Other reviewed approaches include feature engineering with clustering, graph-based classification, and regression-based estimation of non-technical losses.
- Feature selection is applied to customer master and consumption data, while shape factors capture lunch-time, night, and weekend consumption effects for K-means clustering.
- Optimum path forests classify feature-space graph nodes according to their strongest connected training-sample prototypes.
- Optimum path forests train faster but predict more slowly than parametric methods because they do not learn parameters.
- For 736 customers, optimum path forests achieve a test accuracy of 0.9021 and outperform SVMs with Gaussian and linear kernels.
- A regression model estimates technical losses from temperature-dependent infrastructure resistance and subtracts them from overall losses to estimate NTLs.
2.8. Summary
The reviewed NTL detection literature spans diverse models and data sets, but inconsistent evaluation and reporting make comparisons difficult.
- The review covers Boolean and fuzzy expert systems, SVMs, neural networks, genetic methods, optimum path forests, and regression methods.
- Data sets range from 30 to 700K customers, while the 1.1M-customer data set is excluded because it reports only variance reduction.
- Accuracy ranges from 0.45 to 0.99 and recall from 0.29 to 1, whereas precision is reported by few publications from 0.51 to 0.85.
- Comparing methods remains challenging because studies use different data sets and often omit fundamental data properties such as meter readings and NTL proportion.
- The table summarizes models, data sets, and performance measures using two-decimal precision.
3. Challenges
The survey identifies recurring challenges in NTL detection, including class imbalance, feature design, noisy labels, covariate shift, scalability, and limited comparability across data sets. These issues constrain reliable evaluation and deployment of detection models.
- 3.1. Class imbalance and evaluation metric: Class imbalance makes accuracy and recall misleading for NTL detection and requires appropriate evaluation measures.With 999 regular and 1 NTL customer, always predicting non-NTL yields 99.9% accuracy, while always predicting NTL yields 100% recall but triggers costly inspections.
- 3.2. Feature engineering: Hand-crafted features strongly affect classifier performance, but most NTL studies report experimentally improved models without strong theoretical feature justification.Features are derived from monthly or smart-meter consumption profiles and customer master data, whose results are not easily interchangeable.
- 3.3. Data quality: Inspection labels can be incorrect because fraud may be concealed, technicians may be influenced, or data may be processed incorrectly.Most NTL research uses supervised methods, while incorrect training labels are only rarely reported and label-noise handling is less commonly addressed.
- 3.4. Covariate shift: Covariate shift occurs when inspected training customers differ from production customers, causing learned bias and less reliable NTL predictions.Bias can arise from neighborhood-focused inspections and may also involve meter type or connection type; the survey proposes quantifying and reducing this shift.
- 3.5. Scalability: Most reviewed NTL detection methods do not scale computationally to data sets containing hundreds of thousands or millions of customers.Some approaches are feasible only up to tens of thousands of customers, whereas large-scale studies use linear-kernel SVMs or genetic algorithms.
- 3.6. Benchmarking and comparability: Comparisons across NTL detection methods are difficult because studies use different data sets and often omit fundamental data properties.The survey calls for joint research efforts addressing benchmarking and comparability of NTL detection systems.
4. Suggested Methodology
The suggested methodology targets six major challenges in NTL detection, including class imbalance, feature description, inspection bias, scalability, and comparability. It proposes evaluation, feature-learning, bias-correction, and data-set strategies to address them.
- 4.1. Handling class imbalance and evaluation metric: AUC is proposed as a suitable single measure for imbalanced NTL classification because it evaluates the trade-off between true and false positive rates.The goal is to reduce costly false positives while finding as many NTL occurrences as possible.
- 4.1. Handling class imbalance and evaluation metric: Models should be evaluated across different NTL proportions because fixed proportions do not provide imbalance-independent results.The paper reports that precision usually grows linearly with NTL proportion, while recall behaves non-linearly and F1 is unreliable.
- 4.2. Feature description: Temporal processes and neighborhood information are proposed to model self-exciting NTL events and interactions among customers.The methodology considers optimizing the number of temporal processes, since one process for all customers could underfit.
- 4.2. Feature description: Deep learning is suggested for self-learning features from raw consumption data, addressing limitations in manually describing temporal NTL patterns.Most existing research primarily uses features derived from traditional or smart-meter consumption time series, whose results are not easily interchangeable.
- 4.3. Correction of spatial bias: Spatial point processes are suggested to reduce spatial covariate shift caused by inspection samples concentrated in particular neighborhoods.The proposed correction aims to produce more reliable NTL predictors from biased inspection results.
- 4.4. Big Data implementation and 4.5. Creation of a publicly available real-world data set: The paper recommends publicly available real-world data sets and multi-criteria customer ranking to improve model comparison and reflect inspection costs and potential revenue.Data acquisition should balance the amount of data against its cost, while evaluation criteria can include inspection costs and possible revenue increases.
5. Conclusion
The survey reviews AI-based NTL detection and concludes that machine learning methods, especially support vector machines and neural networks, commonly outperform expert systems. It identifies six open challenges whose resolution is needed for sound, comparable, replicable, and scalable results.
- 5. Conclusion: Support vector machines and neural networks are the most commonly used methods and outperform expert systems in most settings.They are typically applied to consumption-profile features and customer master-data features.
- 5. Conclusion: The survey identifies six open challenges: class imbalance and metrics, feature description, incorrect inspection results, covariate shift, Big Data scalability, and cross-method comparability.The challenges span data quality, modeling, evaluation, and implementation.
- 5. Conclusion: Addressing these challenges is presented as necessary to advance NTL detection and produce sound, assessable, understandable, replicable, and scalable research results.The authors report that their current research has started addressing the challenges through the suggested methodology.