Estimation of wind turbine responses with attention-based neural network incorporating environmental uncertainties

The performance of a wind turbine is impacted by various environmental factors and their interactions, such as wind speed, wave load, and temperature fluctuation. In order to evaluate the structural design or the fatigue damage of wind turbines, the uncertainties among environmental factors should be considered in the evaluation framework. However, traditional neural network evaluates wind turbine performances in a deterministic way and the confidence level for prediction results is hard to determine. In the current study, a prior-constrained attention-based neural network (PRANN) model is proposed for wind turbine response evaluations incorporating the uncertainties among environmental factors and their interactions. The proposed PRANN model accounts for the environmental factors separately and constrains their interactions with a conditional autoregressive model to enhance the safety and reliability of wind turbines. To demonstrate the applicability of the PRANN model, a wind turbine benchmark model is used to investigate the structural response prediction performance as well as the fatigue estimation capability of the proposed model. The results show that the PRANN model has high prediction confidence with different recurrent neural network cells. Meanwhile, the adjacent range for the conditional autoregressive model can further improve the prediction confidence but decrease the prediction accuracy. More importantly, the prediction results from the PRANN model can be represented by the confidence interval so that engineers can use the output results to design or evaluate wind turbines at the desired confidence level. The versatility of the PRANN framework allows for its broad application across a wide variety of structures and systems in engineering fields.