Estimation of probability distribution of long-term fatigue damage on wind turbine tower using residual neural network

Abstract

Fatigue is one of the most significant failure modes in structural and mechanical design. As for wind engineering, the fatigue issue on wind turbine towers is more critical since it concerns not only the structural safety but also the power production of the wind turbine. On the one hand, due to the complex fluid–solid interaction and the sophisticated control system of the wind turbine, it is impossible to predict the probability of fatigue-induced failure on wind turbine towers by analytical method. On the other hand, the structural reliability methods based on a numerical approach are usually time-consuming. To overcome these drawbacks, this work firstly proposes a probabilistic fatigue analysis framework to estimate the fatigue damage of wind turbine tower based on numerical simulations. Then, to reduce the computational cost of numerical approach, a residual neural network (ResNet)-assisted fatigue estimation approach is designed for the assessment of long-term fatigue loads under the proposed probabilistic fatigue analysis framework. The proposed probabilistic fatigue analysis framework estimates the cumulative fatigue damage on the cross-section of wind towers in a probabilistic pattern. The designed surrogate-assisted approach learns a model to approximate the relationship between wind speed and the fatigue damage. Then, this surrogate model can be used to predict fatigue damage under different wind speed so that a large number of simulation can be replaced by model prediction. Consequently, the efficiency of the proposed probabilistic fatigue analysis method can be significantly improved. Our proposed method is validated by numerical studies with a state-of-the-art wind turbine and has been applied in a wind turbine design with real-world wind loads.