Strain virtual sensor for offshore wind turbine jacket supports: A time series transformer approach validated with Alpha Ventus wind farm data

Abstract

Monitoring the structural health of offshore wind turbines in harsh marine environments presents significant challenges due to physical sensor limitations, such as susceptibility to high salinity, strong underwater currents, and the resulting durability and maintenance issues. These challenges increase the need for robust and cost-effective alternatives. Motivated by the urgent demand for sustainable and reliable solutions in offshore wind energy, this study introduces a virtual strain sensor based on the PatchTST architecture, a time series transformer designed to capture local and global patterns in multivariate data. By addressing the limitations of conventional physical sensors, the proposed approach uses advanced deep learning techniques to estimate strain accurately in wind turbine jacket supports. The method involves pre-training the model with self-supervised learning and fine-tuning it using real-world strain data from the Alpha Ventus offshore wind farm. Extensive evaluations using statistical metrics (Pearson correlation, root mean square error) and mechanics-based metrics (normalized fatigue damage error and equivalent strain range) demonstrate the virtual sensor’s superior predictive capabilities. Key results reveal that the model outperforms conventional approaches in reliability, adaptability, and computational efficiency, offering an innovative tool for improving structural health monitoring systems. These outcomes support the deployment of virtual sensors as a durable, cost-effective alternative to physical sensors, contributing to enhanced sustainability and efficiency in offshore wind energy systems.

Topics & Concepts