Interpretable prediction of floating offshore wind turbine dynamic Responses: An attention-based deep learning approach

This study develops, for the first time, an interpretable prediction framework for Floating Offshore Wind Turbine (FOWT) dynamic responses, focusing on mooring line tension and tower-top acceleration. The architecture combines a simplified attention mechanism with a Multi-Layer Perceptron (MLP), using a sliding window approach on time-series data generated from OpenFAST simulations of a 22 MW IEA reference turbine. The simulations represent diverse operational conditions, including normal operation and parked scenarios, with varying wind and wave environments. The model achieves high predictive accuracy, essential for reliable interpretation of the attention mechanism. Also, the analysis reveals that the model attends to different time scales: 2.5 s for mooring tension and 1.25 s for tower-top acceleration. Key input features are consistently prioritized, with their relative importance dynamically adjusted based on the operational state. Mooring tension prediction relies heavily on tower-base forces, lower-tower accelerations, and platform motions, whereas tower-top acceleration is primarily influenced by tower-top moments, upper-tower forces and accelerations, and rotor thrust. Validation using standard MLPs confirms that the high-attention features (7) provide better predictions than low-attention features (12). This interpretable model offers potential insights for ocean engineering, including guiding sensor placement, informing structural health monitoring, and contributing to control system design. • We developed an interpretable attention-based model for FOWTs. • Attention-based model reveals key inputs & temporal dependencies across conditions. • Validation confirms attention model identifies key dynamic response drivers.