Banca de QUALIFICAÇÃO: NATHAN SOUZA DE OLIVEIRA

Uma banca de QUALIFICAÇÃO de MESTRADO foi cadastrada pelo programa.
STUDENT : NATHAN SOUZA DE OLIVEIRA
DATE: 15/06/2026
TIME: 09:00
LOCAL: Sala Virtual
TITLE:

Fraud detection in electrical energy consumption via time series imaging and convolutional neural networks

KEY WORDS:

Multivariate time series classification; electricity fraud detection; time series imaging; Recurrence Plot; convolutional neural networks.

PAGES: 53
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUMMARY:

Electricity consumption fraud by consumer units is a relevant technical and economic challenge for power distribution utilities, directly affecting revenue, the tariffs paid by regular consumers, and the operational safety of the grid. This dissertation proposes a fraud detection methodology formulated as a Multivariate Time Series Classification (MTSC) problem, combining time series imaging techniques and Convolutional Neural Networks (CNNs). The profile of each consumer unit is represented by multiple time series, encoded as volumes of continuous, threshold-free Recurrence Plots, in order to preserve the metric structure of the underlying phase-space trajectories. These volumes feed a multi-branch CNN architecture in which each branch extracts specific representations that are subsequently fused and classified by a shared head. Training incorporates specific strategies to address the severe class imbalance, and model selection is driven by a metric that reflects the asymmetry between the operational costs of false positives and false negatives. Experiments carried out on real-world data from a power distribution utility indicate that the representation based on volumes of continuous Recurrence Plots, combined with the multi-branch architecture, effectively captures discriminative patterns in fraudulent consumption profiles, yielding promising results in the evaluated scenario.

COMMITTEE MEMBERS:
Presidente - 347628 - ADRIAO DUARTE DORIA NETO
Interno - 2579664 - ALLAN DE MEDEIROS MARTINS
Externo ao Programa - 1323598 - DANIEL LOPES MARTINS - UFRN