Diagnóstico automático de desbalanceamento em rotores de aerogeradores utilizando técnicas de inteligência artificial

Abstract

Policies aimed at reducing carbon dioxide emissions receive widespread promotion in environmentally focused campaigns, leading to a significant surge in renewable energy generation. To enhance the wind energy sector’s appeal and profitability, research interest in developing methods for monitoring wind turbine conditions has been steadily growing. This study proposes a solution for detecting rotor imbalances in wind turbines resulting from blade mass or aerodynamic imbalances. Developing this solution required an investigation into which operational signals of the wind turbine are affected when a rotor imbalance occurs. This investigation relied on results obtained through simulation software and analyses of mathematical models of wind turbines. The OpenFAST software was employed for simulations, comprising structural, aerodynamic, electrical, and control models of the wind turbine. To generate wind time series, the TurbSim software, also developed by NREL, was used. It is important to mention that OpenFAST allows integration with Simulink, a Matlab tool, where control loops for maximum power point tracking (MPPT) and power limitation regions were developed. To manage and automate simulations and create a database, a Python application was developed to manipulate OpenFAST and TurbSim files. This application conducts simulations under various wind scenarios and rotor imbalance conditions. Analysis of simulations and mathematical models of the wind turbine revealed that mass and aerodynamic imbalances affect translational accelerations of the nacelle’s inertial units. The nacelle’s translational acceleration in the ys axis direction is more significantly disturbed by aerodynamic imbalance, while acceleration along the xs axis is similarly affected by both mass and aerodynamic imbalances. For automated rotor imbalance detection, two artificial intelligence techniques were employed for comparison. The first technique involved statistical descriptors of accelerations and wind speed, which were used to generate images. These images are fed into a convolutional neural network (CNN) for training and prediction. The second method utilized accelerations in the frequency domain, focusing on frequencies near 1p. In this case, a support vector machine (SVM) was used for training and prediction of rotor conditions. Additionally, this study introduces an innovative approach by incorporating the use of Generative Adversarial Networks (GANs) for domain adaptation. This strategy aims to make the rotor imbalance detection method highly adaptable to different wind turbine models. The significant advantage of GANs lies in their capacity to adapt the detection model without requiring a complete algorithm overhaul, making it efficient and practical for different models. Finally, a practical solution for implementing the technique in wind farms is proposed by integrating it into a management system that enables continuous monitoring, real-time detection, and predictive maintenance, contributing to optimized performance and integrity of industrial-scale wind farms.