A New Two‐Stage Probabilistic Remaining Useful Life Prediction Method for Wind Turbines

Accurate and reliable modeling of the degradation process is crucial for developing preventive maintenance strategies for wind turbines (WTs). However, the degradation data of WTs exhibit complex patterns and are often interfered with by noise, making it challenging to extract meaningful features from the data. Additionally, existing health indicators used in determining the first prediction time (FPT) of the remaining useful life (RUL) often fail to reflect the true health state of WT components, resulting in inaccurate starting points for RUL predictions. Furthermore, current methods for quantifying uncertainties in sensor data and predictive models typically require extensive degradation data and significant computational resources. To address these issues, this paper proposes a multihead long short‐term memory neural network, which decouples complex time series data to extract prior sequential features and employs a parallel architecture to improve prediction accuracy. Building upon this, a two‐stage probabilistic RUL prediction method is proposed. In the first stage, the FPT is determined based on a normal behavior model. In the second stage, the uncertainty of the RUL is quantified by integrating the linear regression and the kernel density estimation strategies, which provide probabilistic RUL predictions. The effectiveness of the proposed method is validated by two case studies using a publicly available dataset and real WT data, respectively. The results demonstrate that the proposed method can provide accurate and robust RUL prediction results.