Session: 41-05 Artificial Intelligence Applied to Wind Energy

Paper Number: 152034

Multivariate Time-Series Modelling for Wind Turbine Subsystem Reliability Prediction 

Wind turbine reliability monitoring and prediction are essential for the early detection of potential failures, enabling proactive maintenance, optimizing turbine performance and component efficiency, and minimizing the levelized cost of energy. To date, many utility-scale wind turbines worldwide are approaching the midpoint or end of their operational lifespan, underscoring the increased importance of reliability monitoring and prediction. The existing investigations into the statistical analysis of component failures do not fully account for dynamic changes or the need to prevent costly breakdowns. Real-time health monitoring of turbine signals is inadequate for proactive maintenance or repairs of critical components, such as blades or gearboxes. Consequently, we propose a physics-informed wind turbine reliability probabilistic prediction (PI-WTRPP) model to facilitate quarterly forecasts of the occurrence, duration, and type of potential failures or probability of downtime for major turbine components, including the blade, hub, gearbox, bearings, generator, and electrical subsystem. The PI-WTRPP model is developed using a multivariate time-series transformer (MTST) architecture, which integrates neural ordinary differential equations (Neural ODEs) within a multi-head attention mechanism to capture the temporal correlations of turbine operational features and dynamic behaviors of the aforementioned major components. The physics-informed equations are embedded into the model as loss functions to enhance the generalization for various turbine operating scenarios. The proposed model was systematically developed and evaluated using ten years of operational data from a utility-scale horizontal-axis wind turbine (2.5 MW) located in the Midwest region of the United States. Data cleaning was performed prior to model training and testing to eliminate downtime events caused by factors such as low wind inflow below the turbine's cut-in wind speed and power system curtailment, retaining only those downtime events attributed to turbine component failures or abnormal performance. Preliminary results indicate that the proposed PI-WTRPP model can reliably predict potential downtime within the next three months and identify the type of abnormal performance or failures. In comparison to existing statistical methods and basic machine-learning models, the PI-WTRPP model achieves optimal performance across various evaluation metrics, highlighting the model's robustness. Quarterly forecasts of the occurrence, duration, and type of potential failures, as well as the probabilities of downtime for major turbine components, can enhance decision-making in wind turbine maintenance and component management, ultimately reducing maintenance costs and extending the turbines' lifespan.

Presenting Author: Fuhao Chen University of Colorado Denver

Presenting Author Biography: Fuhao Chen is a PhD student in the Department of Mechanical Engineering at the University of Colorado Denver | Anschutz Medical Campus, supervised by Prof. Linyue Gao. He earned his bachelor's (2021) and master's (2024) in Renewable and Clean Energy from North China Electric Power University and was a visiting student at KTH Royal Institute of Technology from 2022 to 2023. His research focuses on wind power forecasting, wind farm modeling, data-driven modeling, and deep learning.

Authors:

Fuhao Chen University of Colorado Denver
Linyue Gao University of Colorado Denver