Análise, avaliação e validação do uso de técnicas de aprendizado de máquina para detecção de falhas em turbinas eólicas do tipo PMSG - Gerador síncrono de ímãs permanentes

Abstract

The implementation of sustainable energy sources is constantly expanding, with particular emphasis on wind power, which plays a significant role in meeting carbon emission reduc tion targets and diversifying energy supply. Wind turbines are highly complex systems, and the high costs associated with operation and maintenance have spurred the search for increasingly effective procedures to reduce them. This study proposes that deviati ons from expected system behavior in wind turbines be detected and classified as faults using a data-driven approach, leveraging operational data from various sensors installed in direct-drive PMSG (Permanent Magnet Synchronous Generator) wind turbines. This data is acquired through the Supervisory Control and Data Acquisition (SCADA) system provided by ELETROBRAS, located in Brazil, specifically in the state of Bahia, Casa Nova. To achieve fault detection and capture information about the turbine’s condition, a machine learning methodology is employed. The CFS (Correlation-based Feature Se lection) and Random Forest methods are used to select the most relevant attributes from the dataset. Subsequently, two Multi-layer Perceptron (MLP) neural networks are con figured, validated, and evaluated, using the most relevant attributes selected by the two aforementioned methods as input to classify whether wind turbines exhibit fault behavior or operate normally. The results demonstrate that the strategy of selecting the most relevant attributes for the MLP neural network for fault classification resulted in a high rate of fault detection in wind turbines. The CFS-MLP neural network showed an excel lent percentage of instances correctly classified, at 98.73%, with an average precision of 98.9% for fault occurrences and 98.5% for correct classification of normal turbine opera tion, using the dataset. Meanwhile, the RF-MLP neural network exhibited a percentage of instances correctly classified at 97.88%, with an average precision of 98.4% for fault occurrences and 97.4% for correct classification of normal turbine operation. However, assessing performance alone is not sufficient; it is also important to consider the reduction of the number of selected sensors (attributes).