Multi-task learning long short-term memory model to emulate wind turbine blade dynamics

Abstract

The substantial computational expense associated with the dynamic analysis of wind turbines prohibits efficient design evaluations and site-specific performance predictions. This research explores the effectiveness of principal component analysis and discrete cosine transform dimensionality reduction methods to identify key spatial and temporal patterns in a wind field, which are subsequently used by a long short-term memory (LSTM) algorithm to model the wind turbine responses. This study strikes a balance between prediction accuracy and training data requirements by employing an efficient feature selection technique and a multi-stage modelling approach that incrementally learns the information about the target variable. Furthermore, a multi-task learning strategy is adopted, allowing the LSTM model to predict multiple target variables at once, thus removing the necessity for separate models for each target variable. This method alleviates the computational cost of dynamic analysis of a wind turbine by addressing the challenges introduced by high-dimensional wind fields and time-consuming numerical integration processes. The findings show that this comprehensive approach significantly reduces computational cost while maintaining accuracy across all target variables, thereby facilitating design feasibility assessments and site-specific studies of wind turbines.