Abstract
Electricity theft is more than a technical irregularity—it undermines the financial sustainability
and reliability of utilities across developing economies. In Kinshasa, Democratic Republic of
Congo, the problem is intensified by fragile infrastructure, entrenched corruption, and the limited
reach of conventional detection methods. To address these systemic barriers, this dissertation
introduces a hybrid machine learning framework that integrates behavioural analytics, recursive
feature selection, ensemble learning, and a corruption‑aware investigation mechanism. Together,
these elements are designed to confront systemic barriers and offer a more resilient approach to
detecting and prioritising electricity theft in fragile grids.
At its core, the framework builds a multi‑stage detection pipeline. Domain‑specific behavioural
features—such as the Temporal Gap Deviation Score (TGDS), Efficiency Variability Index (EVI),
and Time Since Last Transaction (TSLT)—are engineered from consumer usage data. These are
refined through recursive feature selection to identify the most reliable inputs, which then feed into
supervised ensemble models like Random Forest, Gradient Boosting, and Logistic Regression.
The models achieve exceptionally high accuracy (F1‑score: 99.95%), while an unsupervised
ensemble approach delivers precise anomaly detection (100% precision at a 5% threshold).
The pivotal innovation is the Bayesian Recursive Ranking (BRR) mechanism, which dynamically
prioritises suspected theft cases through credibility‑weighted belief tracking and adaptive case
selection. To reflect fragile governance realities, the system incorporates an honesty ratio
calibrated against the DRC’s Corruption Perceptions Index, adjusting the reliability of investigator
feedback.
Simulation studies grounded in real world consumer data confirm the robustness of the framework.
BRR consistently converges to accurate beliefs despite noisy and biased inputs, achieving over
97% success within shallow recursion. It sustains stable audit depth across disturbance levels,
adapts to shifting behavioural patterns, and maintains detection efficiency with modest resource
investment. Importantly, its resilience comes not from brute force, but from adaptive
recalibration—preserving accuracy and investigative integrity even when corruption distorts the
evidence...