| description abstract | Abstract. With the rapid development of the offshore wind power industry, ensuring the stability and safety of wind turbines has become increasingly important. As a critical component, the temperature of wind turbine bearings serves as a key indicator for assessing system health and predicting potential failures. However, conventional monitoring methods often struggle to cope with the complex and dynamic offshore environment. To address this challenge, this study proposes an integrated approach that combines time-series enhanced Bayesian random forest imputation, a transformer model with adaptive sparse attention, and bootstrap resampling. The imputation method leverages Bayesian optimization to adaptively tune hyperparameters and incorporates the interaction between temporal features of the temperature sequence and environmental covariates, improving the physical consistency and robustness of the results through iterative refinement. The adaptive sparse transformer effectively captures key temporal dependencies in the bearing temperature series, enhancing the model's ability to learn complex data patterns. Meanwhile, the bootstrap resampling technique dynamically generates confidence intervals, quantifying prediction uncertainty and providing a more reliable foundation for anomaly detection. The proposed method is validated using supervisory control and data acquisition data from offshore wind turbines located off the West African coast. Experimental results demonstrate excellent performance in both predictive accuracy and anomaly detection, achieving an R2 of 0.9954 and a root mean squared error of 1.3545. These results confirm the method's strong potential for wind turbine bearing temperature forecasting, offering scientific support for intelligent operation and maintenance strategies, as well as practical solutions for early fault warning and resource optimization. | |
