|
Abstract
|
There is a growing concern about the high degree of Non-Technical Losses (NTL) in developing countries especially sub-saharan Africa. Whereas several studies have employed artificial intelligence (AI) to analyze and detect NTL, these studies focus on customer consumption data. To the best of our knowledge, no research has empirically explored the activities of electricity distribution staff as possible contributors to NTL. Furthermore, limited or no research has studied the decision rationale influencing how AI models interpret the significance of features in predicting NTL. This paper explores the nexus of electricity consumers, staff and NTL by analyzing a combined dataset of customer consumption and staff activities data. A novel deep-learning architecture called NTLCONVNET was developed and compared with 5 ensemble learning techniques for predicting NTL. To actualize this, a dataset containing 12 input features and one labeled output was curated and 5-fold cross-validation was used during the model training to minimize bias. SHAP algorithm kernel and tree-based explainers were employed for model interpretability of the deep-learning and ensemble-learning models respectively, and a novel holistic ranking approach was used to rank the features model predictive significance. The results show that the NTLCONVNET significantly surpasses all other approaches on both weighted and macro averages of the performance metrics at a p-value p < 0.05 . The results show that 4 out of the top 6 most influential features for model prediction are staff related. This study, suggests a policy outcome of introducing human resource metrics into NTL reduction strategies.
|