Hybrid surrogate input load estimation model in offshore wind turbines using transfer learning and multitask learning

Accurate input load estimation is useful for proactive maintenance, design optimization, and digital twining of offshore wind turbines. Traditional physics-based approaches for input load estimation, while accurate, face challenges due to high computational demands, limited availability of experimental data, and challenges with direct measurement of input loads. This paper introduces a novel hybrid surrogate model for real-time input load estimation that integrates data-driven and physics-based methodologies to address these limitations. The proposed framework focuses on training a neural network-based regression model that can predict the input load time history using the dynamic structural response (accelerometers and strain measurements) as input. The model is optimally designed to overcome the lack of experimental data by following a transfer learning strategy: the model is initially trained on synthetic data generated from a physics-based model and then fine-tuned using experimental data. Additionally, a multitasking approach is implemented to extend the applicability of the model to different regions of operation. Finally, a dropout Montecarlo strategy is proposed to quantify the uncertainty of the regression model. This hybrid surrogate model demonstrates superior computational efficiency and accuracy in input load estimation compared to conventional physics-based models, and it offers several practical benefits. It is scalable, does not necessitate continual recalibration, and exhibits robustness across diverse datasets and operational conditions. The model's potential for real-time applications, such as digital twin technology, underscores its contribution to advancing maintenance strategies and design improvements in offshore wind turbine engineering.