Enhancing long-term motion prediction for floating energy platforms: A physics-informed framework multi-scale recalibrated fusion transformer model

To address the persistent challenge of long-term motion prediction accuracy for offshore floating energy platforms, this study proposes a novel multi-scale recalibrated fusion Transformer (MRF-Transformer) model. This model integrates physics-informed multi-scale convolutional fusion to capture critical ocean dynamics, an environment-adaptive gating mechanism, and dual-residual structures, collectively overcoming phase drift and error accumulation limitations of conventional models during extended horizons. Validation with field data from the South China Sea's Lingshui 17-2 gas field platform demonstrates transformative performance: At second-scale prediction, mean absolute error (MAE) reaches 0.0128°—a 22.9% reduction versus Transformer—while R2 achieves 0.9364. Under typhoon conditions, roll prediction RMSE reduces to 0.0399°, representing a 36.9% improvement. Crucially, for 15-min horizons, it maintains R2 = 0.2233, significantly outperforming long short-term memory (LSTM) which degraded to −0.0565, thereby extending safe operational decision windows. By optimizing motion compensation and equipment scheduling, MRF-Transformer enhances energy efficiency and reduces annual CO2 emissions. The framework generalizes to floating wind turbines and wave energy converters, advancing low-carbon offshore renewable energy development.