Artificial Intelligence for Management of Non-Technical Losses in Electricity Distribution Networks in Sub-Saharan Africa: A Case Study of Nigeria

Sub-Saharan African countries such as Nigeria have consistently faced the problem of non-technical losses (NTL) in their electricity distribution networks, mainly caused by energy theft and operational deficiencies. These losses undermine the financial and operational viability of utility companies, leading to unreliable electricity supply with adverse socio-economic consequences. Previous studies on NTL have used artificial intelligence (AI) to analyse electricity consumption patterns to detect anomalies, hence they had primarily focused on the electricity consumer as the major cause of NTL. The author is among the pioneer researchers to conduct preliminary research which indicates that there is increasing anecdotal evidence which suggests that the activities and involvement of utility company staff significantly contribute to NTL. Yet, to the best of the author’s knowledge, no empirical studies have been undertaken to systematically investigate this phenomenon. Therefore, existing literature in this field suffers from an "omitted variable defect" as it fails to account for a comprehensive range of factors necessary for addressing the problem of NTL effectively. This limitation hinders the ability to develop holistic solutions because critical variables, particularly those related to utility staff activities, are often overlooked in current analyses. This thesis represents the first empirical investigation into the impact of utility staff activities on NTL by analysing staff-related data alongside customer consumption and other customer-related data. Unlike previous studies that focused on customers, this thesis presents a novel approach to reducing NTL by developing a hybrid machine learning (ML) model called 4XCNNBoost to analyse a combined dataset of customer consumption and staff operations. The model consists of one dimensional convolutional neural network (1D CNN) and extreme gradient boosting (XGBoost) model, combining the strengths of deep learning algorithms in capturing complex patterns in large datasets with superior classification accuracy of gradient boosting algorithms. The dataset used for this study consists of 24-month record post-paid customer consumption and staff operational activities of an electricity distribution company in Nigeria. Three optimisation techniques namely, Particle Swarm, Bayesian, and GridSearchCV were used for model hyperparameter optimization with GridSearchCV achieving the average highest accuracy of 0.87 for the CNNs and 0.98 for the XGBoost metamodel. The optimized 1D CNN models also show recall range of 0.29–0.94, and macro F1-score (0.53–0.92), with GridSearchCV-tuned CNN2 achieving peak accuracy (0.94) and recall (0.94). A novel features (variables) ranking framework was developed using SHapley Additive exPlanations (SHAP) explainability algorithm to provide insights into the relative significance of customer and staff-related variables in causing NTL. The result shows a ranking score range of 0.0192 to 0.2500 with the collection_index ranking highest across all input models with a ranking score of 0.2500.   The result of the experiment reveals that over 57% of top causes of NTL are staff-related.  This experimental setup is repeated using 4 other machine learning models namely Random Forest, Support Vector Machine, Decision Tree and Logistic Regression and the result also show that over 57% of top causes of NTL are staff-related. The results also show that the 4XCNNBoost model outperforms the four models with an accuracy of 0.98, precision of 0.99, recall of 0.97 and f1 score of 0.98. To further validate the robustness of this findings, the experiment was repeated using a consolidated national electricity distribution dataset in Nigeria, and the result reveal that over 80% of major causes of NTL are staff-related, hence is consistent with the result of company-specific dataset. Furthermore, given that the analysis identifies staff-related factors as major sources of NTL, this study extends its scope by drawing a parallel between NTL and operational risk by highlighting how human behaviours and technological vulnerabilities contribute to operational inefficiencies. Consequently, these causes of NTL are mapped into operational risk cells using the BASEL II and III risk definitions, while the Three Lines of Defence (3LoD) model is used to formulate NTL mitigation strategy. This thesis makes conceptual, methodological and practical contributions to the domain of electricity distribution networks and engineering management. First, by highlighting a significant omitted variable defect in existing literature, this work opens up a new conceptual and methodological perspective for understanding the broader drivers of NTL. This challenges the previously held empirical assumption that customers are the primary contributors to NTL, offering a more comprehensive viewpoint on the underlying causes. This study introduces a novel approach of integrating staff-related variables into the empirical analysis of NTL, a dimension hitherto overlooked in existing literature. Furthermore, this research contributes to the discourse of advancing the use of hybrid machine learning models by demonstrating the effectiveness of such models, hence highlighting the possibility for even more sophisticated models capable of learning complex relationships in large heterogeneous datasets. Also, to the best of the author’s knowledge, it is the first study to apply the 3LoD model to electricity distribution such that utility companies to leverage the well-established Operational Risk Management frameworks to enhance operational resilience and mitigate staff-related factors driving NTL. The study also makes a population group contribution by expanding the geographical scope of NTL research to target the under-researched region of Sub-Saharan Africa. Consequently, it presents a regional contextualization of a global issue by providing insights into the specific nuances and peculiarities of NTL in this region. The findings of this research have theoretical, practical, and policy implications that improve efficiency and provide a foundation for further research to address the challenges of NTL. The novel approach of analysing a combined staff and customer dataset expands the theoretical scope of NTL analysis for a more holistic insights. The study also provides a practical tool for developing targeted interventions to improve the efficiency of electricity distribution networks.  In terms of policy implications, the knowledge generated from this research will significantly help regulators and other policy stakeholders to better understand the systemic factors of NTL, and design more practical regulatory frameworks. This is especially relevant in contexts where the sector faces challenges of inefficiency or corruption. The findings of this research are not limited to electricity distribution but hold broader implications by finding relevance in other infrastructure-based engineering domains where resource allocation, revenue optimisation, and customer/staff-centric approaches are critical. The research focus on AI, operational risk management and policy simulation demonstrates a pathway for operational sustainability using data-driven and cross-disciplinary approaches to solving infrastructure challenges.