Predicting Wind Turbine Blade Tip Deformation With Long Short-Term Memory (LSTM) Models

Abstract

Driven by the challenges in measuring blade deformations, this study presents a novel machine learning methodology to predict blade tip deformation using inflow wind data and operational parameters. Using a long short-term memory (LSTM) model and a novel feature selection approach based on mutual information and recursive feature addition, this study presents a robust frame- work for multivariate time series prediction. The developed model offers significant computational cost reductions compared to full-dynamic simulations and also allows virtual sensing. This work empowers efficient and reliable wind turbine operation by providing an accurate and computationally efficient blade response prediction tool that can assist in improved wind turbine management, site-specific analysis and fatigue assessment.