15Jun2026
15:00 Doctoral defense UNICAUCA (Colombia)
Topic on
Network Slicing in 5G for communication network management in power substations
Student
Edwin Ferney Castillo Quintero
Advisor / Teacher
Nelson Luis Saldanha da Fonseca and Oscar Mauricio Caicedo Rendon (joint supervision - UNICAUCA)
Brief summary
Smart Grid 2.0 (SG 2.0) integrates advanced communications, Machine Learning (ML), and Advanced Metering Infrastructure (AMI) to enhance the control, protection, communication, and monitoring capabilities of electric power systems. Electricity Theft Detection (ETD) is fundamental to SG 2.0 because theft generates economic losses exceeding US$25 billion annually worldwide, prevents Distribution System Operators (DSOs) from recovering unpaid energy costs, and causes voltage imbalances, transformer overloads, and increased operating costs. Existing electricity theft detection solutions utilize large-scale consumption data collected by AMI, along with centralized ML and Federated Learning (FL), to accurately detect electricity theft. However, these approaches often neglect the need to simultaneously achieve high detection performance, low resource consumption, and explainable decision-making processes. This limitation is particularly relevant in SG 2.0, where the results of the ETD (Environmental, Technical and Disaster Risk Assessment) can lead to inspections, penalties, or service interruptions, requiring auditable and transparent decisions.
This thesis develops an approach based on eXplainable Artificial Intelligence (XAI) and FL for an effective and explainable ETD in SG 2.0 enabled by Network Slicing. Specifically, it seeks to: (i) design an XAI and FL-based approach to explain and optimize ETD; (ii) prototype mechanisms that improve the effectiveness and explainability of ETD; and (iii) evaluate the approach in terms of learning performance, communication and computation overhead, and explainability. This approach comprises two main components. First, an explainable approach to customer selection, called xCS, provides effective, explainable, and reliable customer selection during federated training. xCS extends traditional FL through Ranking and Quartile policies, supported by algorithms for applying these policies and calculating the contribution value. Customer contributions are estimated using game theory indices, including Banzhaf (BZ), Least Core (LC), Johnston (JT), and Shapley Value (SV), increasing the explainability and reliability of customer selection. Secondly, an explainable approach to ETD, called xETD, combines FL, XAI, and model compression techniques (i.e., pruning and quantization) to achieve high detection efficiency with low resource consumption. xETD extends FL through two XAI-based policies: Adaptive Uplink Compression (AdapUp) and Static Downlink Compression (StatDown).
Evaluations of xCS under different levels of data heterogeneity have demonstrated its effectiveness and robustness. The results indicate that BZ, LC, and JT are recommended for large-scale or time-sensitive scenarios where reducing training duration is critical, while SV remains preferable when the primary goal is to maximize learning performance. Robustness analyses showed that policies based on BZ produced more stable contribution values under data perturbations than alternatives based on SV, LC, and JT. These findings demonstrate that BZ, LC, and JT offer lightweight and efficient alternatives to SV, while maintaining comparable learning performance. Evaluations of xETD have demonstrated its effectiveness and reduction in resource consumption. The results show that xETD reduces communication and computation overhead, as well as convergence time, incurring only marginal reductions in accuracy and F1-score compared to existing FL-based ETD approaches. Taken together, the results demonstrate that the combined use of xCS and xETD constitutes a promising solution for effective, explainable, and reliable electricity theft detection in SG 2.0.
Examination Board
Headlines:
| Nelson Luis Saldanha da Fonseca | IC / UNICAMP |
| Oscar Mauricio Caicedo Rendon | DTm/UNICAUCA |
| Gustavo Adolfo Ramirez Gonzalez | DTm/UNICAUCA |
| Carlos Alberto Astudillo Trujillo | IC / UNICAMP |
| Edmundo Roberto Mauro Madeira | IC / UNICAMP |
| Jeferson Campos Nobre | INF / UFRGS |
Substitutes:
| Não há | --- |