Attention-based machine learning for short-term motion prediction of a floating offshore wind turbine

For over half a century, the forward prediction of random wave series and their induced float motions has remained a challenging problem, hindered by the intricate dynamics of wave–wave and wave–structure interactions. Recent advances in artificial intelligence (AI) and computational power have reignited efforts to transcend the limitations of conventional methodologies. Here, we present a data-driven framework to predict the heave motion of a floating offshore wind turbine using machine learning models trained on full-scale field measurements from the TetraSpar demonstrator. Four machine learning models are evaluated, among which the attention-based encoder–decoder Long Short-Term Memory model provides the best overall performance, achieving reliable predictions for up to two-thirds of a characteristic heave motion period under typical sea states. These results highlight the potential of machine learning methods for short-term motion forecasting and support the development of active control strategies in renewable energy systems, with potential benefits for the economic viability of offshore renewable energy development. • A data-driven framework is developed for phase-resolved heave motion forecasting of a floating offshore wind turbine. • The proposed LSTM model achieves reliable predictions up to two-thirds of a typical heave period. • The model is validated using full-scale field measurements for floating wind energy turbine. • The framework offers potential benefits for safer personnel transfer operations and more efficient control of floating wind turbines.