Fault Diagnosis of Wind Turbine Generators Based on Stacking Integration Algorithm and Adaptive Threshold

Area of Science:

• Renewable Energy Systems
• Machine Learning Applications
• Predictive Maintenance

Background:

• Fault alarm time lag hinders accurate wind turbine generator (WTG) diagnosis.
• Current methods are insufficient for rapid WTG fault detection, increasing operational costs.
• Efficient fault diagnosis is crucial for reliable wind energy production.

Purpose of the Study:

• To develop a novel, fast, and accurate fault diagnosis method for WTGs.
• To reduce operational and maintenance costs associated with WTGs.
• To improve the reliability and efficiency of wind turbine operations.

Main Methods:

• Constructed a stacking ensemble model using LightGBM, XGBoost, and SGDRegressor.
• Employed Bayesian tuning for automatic hyperparameter optimization.
• Applied the Pauta criterion (3σ) and temporal sliding window with fitted residuals for an adaptive threshold method.

Main Results:

• The proposed adaptive threshold method outperformed fixed thresholds in fault diagnosis.
• Achieved significantly earlier alarm times: 1.5 h for GENERATOR, 5.8 h for GENERATOR_BEARING, and 3 h for TRANSFORMER faults.
• Demonstrated the model's accuracy and applicability using R² and RMSE metrics.

Conclusions:

• The developed adaptive threshold method enables faster and more accurate WTG fault diagnosis.
• This approach effectively addresses the challenge of fault alarm time lag.
• The findings contribute to reducing WTG operational costs and enhancing energy system reliability.