An Unsupervised Machine Learning Framework for Fraud and Anomaly Detection in Nigerian Prepaid Electricity Transactions

Prepaid electricity metering is widely adopted in Nigeria to improve revenue collection and reduce customer indebtedness. However, irregularities in transaction records continue to challenge operational reliability and financial transparency. This study presents an unsupervised machine learning framework for detecting anomalies in prepaid electricity transactions using nine months of real-world data. The framework integrates three distinct anomaly detection methods: Isolation Forest, DBSCAN, and a reconstruction-based model using either an Autoencoder or Principal Component Analysis (PCA). These models were combined through a rank-based ensemble scoring system and a majority-vote mechanism to enhance detection of robustness. The dataset includes 23 features spanning customer identifiers, tariff details, and transaction attributes such as energy purchased, payments made, arrears, and VAT. Preprocessing steps involved standardizing column formats, handling missing values, and engineering features such as payment ratios and log-transformed monetary values to improve model sensitivity. Each model independently flagged anomalies, and the ensemble strategy consolidated these outputs to identify high-confidence irregular transactions. The framework uncovered several types of anomalies, including transactions with missing payment and unit values but large arrears repayments, extreme pay-per-unit ratios exceeding operational norms, and VAT entries that deviated significantly from the statutory rate. Spatial analysis revealed concentrated anomalies in specific districts and feeders, suggesting localized vulnerabilities in transaction management and enforcement. Although ground-truth fraud labels were unavailable, the detected anomalies represent statistically significant deviations that warrant further investigation. The results demonstrate that unsupervised models can effectively highlight suspicious patterns without relying on labeled data, offering a scalable approach for utilities to monitor prepaid electricity systems. This methodology supports targeted audits, enhances revenue protection, and contributes to improved regulatory compliance. The study underscores the potential of data-driven techniques in addressing fraud and operational inefficiencies in African energy systems. Future work may incorporate labeled datasets, temporal features, and network-level attributes to refine detection capabilities and expand the scope of analysis.