[Paper Review] Neighborhood Features Help Detecting Electricity Theft in Big Data Sets.

Patrick Glauner, Jorge Augusto Meira+5 more|arXiv (Cornell University)|Jul 4, 2016

Electricity Theft Detection Techniques16 references3 citations

TL;DR

This paper proposes a neighborhood-based feature engineering approach to detect electricity theft in large-scale power distribution networks by analyzing grid-level inspection proportions and non-technical loss (NTL) rates. Using 31 million meter readings across 700K customers, it integrates spatial, temporal, and master data features into four scalable machine learning models, achieving improved detection accuracy across NTL proportions from 1% to 90%.

ABSTRACT

Electricity theft is a major problem around the world in both developed and developing countries and may range up to 40% of the total electricity distributed. More generally, electricity theft belongs to non-technical losses (NTL), which are losses that occur during the distribution of electricity in power grids. In this paper, we build features from the neighborhood of customers. We first split the area in which the customers are located into grids of different sizes. For each grid cell we then compute the proportion of inspected customers and the proportion of NTL found among the inspected customers. We then analyze the distributions of features generated and show why they are useful to predict NTL. In addition, we compute features from the consumption time series of customers. We also use master data features of customers, such as their customer class and voltage of their connection. We compute these features for a Big Data base of 31M meter readings, 700K customers and 400K inspection results. We then use these features to train four machine learning algorithms that are particularly suitable for Big Data sets because of their parallelizable structure: logistic regression, k-nearest neighbors, linear support vector machine and random forest. Using the neighborhood features instead of only analyzing the time series has resulted in appreciable results for Big Data sets for varying NTL proportions of 1%-90%. This work can therefore be deployed to a wide range of different regions around the world.

Motivation & Objective

• To address the widespread problem of non-technical losses (NTL) in electricity distribution, which can reach up to 40% in some regions.
• To improve detection of electricity theft in large-scale datasets by incorporating spatial neighborhood patterns alongside traditional time-series and customer data features.
• To develop a scalable, parallelizable machine learning pipeline suitable for big data environments in diverse global regions.
• To evaluate the effectiveness of neighborhood-based features in enhancing detection performance across varying levels of NTL prevalence.

Proposed method

• Divides the geographical area into grids of varying sizes to aggregate customer data at the neighborhood level.
• Computes two key neighborhood features per grid cell: the proportion of inspected customers and the proportion of non-technical losses (NTL) among inspected customers.
• Generates time series features from customer meter readings and incorporates master data such as customer class and voltage level.
• Applies four parallelizable machine learning algorithms—logistic regression, k-nearest neighbors, linear support vector machine, and random forest—to the feature set.
• Trains models on a Big Data set containing 31 million meter readings, 700,000 customers, and 400,000 inspection results.
• Uses the distributions of neighborhood features to validate their predictive utility for NTL detection.

Experimental results

Research questions

• RQ1 How do neighborhood-level features—such as inspection and NTL proportions—improve electricity theft detection in large-scale datasets?
• RQ2 To what extent do spatial features enhance detection performance compared to time-series and master data features alone?
• RQ3 How does the proposed method perform across different levels of NTL prevalence, from 1% to 90%?
• RQ4 Can the integration of spatial, temporal, and customer-specific features lead to more robust detection in diverse regional contexts?

Key findings

• The inclusion of neighborhood features significantly improved detection performance compared to models using only time-series or master data features.
• The proposed method achieved appreciable results across a wide range of NTL proportions, from 1% to 90%, indicating robustness across varying loss levels.
• The four machine learning models—logistic regression, k-nearest neighbors, linear SVM, and random forest—were successfully applied in a scalable, parallelizable manner on the Big Data set.
• The distributions of neighborhood features were found to be informative and useful for predicting NTL, supporting their integration into detection pipelines.
• The approach is deployable across diverse regions due to its adaptability to varying NTL levels and scalable architecture.
• Spatial features derived from grid-based aggregation provided meaningful context that enhanced model performance beyond individual customer behavior.