Determinação da contribuição harmônica em pontos de acoplamento comum de microrredes: uma abordagem baseada em aprendizado de máquina

• Ano de defesa: 2025
• Autor(a) principal: Mella, Matheus Clementino de
• Orientador(a): Fernandes, Ricardo Augusto Souza
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Engenharia Elétrica - PPGEE
• Palavras-chave em Português: Aprendizado de máquina; Contribuição harmônica; Distorção harmônica; Qualidade da energia elétrica; Sistemas de distribuição de energia
• Palavras-chave em Inglês: Machine learning; Harmonic contribution; Harmonic distortion; Power quality; Power distribution systems
• Área do conhecimento CNPq: ENGENHARIAS::ENGENHARIA ELETRICA::SISTEMAS ELETRICOS DE POTENCIA
• Link de acesso: https://hdl.handle.net/20.500.14289/22094

The increasing integration of nonlinear loads and distributed energy resources in microgrids (MG) has raised the levels of harmonic distortions in distribution feeders. Identifying the origin of these distortions, whether from the MG or the utility’s medium-voltage network, is essential for managing the quality of electrical power. In this context, this project proposed a machine learning-based approach to identify the side responsible for the location of harmonic sources in electrical distribution systems, determining whether the main contribution to the distortions occurs on the utility side or the MG side. For this, Random Forest and XGBoost classifiers were used. The F1-score was adopted as the evaluation metric because it combines precision and recall metrics, providing a balanced view of the model’s performance, especially in scenarios where there is class imbalance. The methodology was validated through simulations on the IEEE 34-bus feeder using the ATP software, with harmonic sources such as 6 and 12-pulse rectifiers, SFC, DC motor drive, and TCR, located on the utility side or the microgrid side, with each simulation considering only one harmonic source. The generated dataset was used to train and validate the models, which achieved F1-scores above 99%, showing high precision in identifying the predominant side in harmonic distortions. However, the F1-score decreased as the distance between the source and the PAC increased due to harmonic distortion attenuation. Nevertheless, both models maintained F1-scores above 97% for distances greater than 40 km, demonstrating the robustness of the models. The entire process, including scenario definition, simulation execution, dataset structuring, and machine learning pipeline implementation, was automated in Python. The Optuna library was used to optimize the models by continuously adjusting the classifiers’ parameters. The methodology proved effective in identifying the side responsible for harmonic distortions, with robust performance even in large-distance scenarios. The results indicate that the approach is useful in the simulated scenario, considering the location of one harmonic load at a time, allowing a clear analysis of the location of harmonic sources in the distribution system.