A high-fidelity physics-constrained deep learning framework for cross-condition wake prediction in tidal current turbines

The prediction of wake characteristics from tidal current turbines (TCTs) is crucial for optimizing array-scale layouts. However, current rapid surrogate models for wake prediction often degrade under variations in inflow distribution. This study proposes a high-fidelity, physics-constrained deep learning framework for reliable cross-condition prediction of TCT wake characteristics. The framework uses a 4D input, comprising inflow velocity and turbulence intensity (TI) at upstream locations, to predict plane-wise wake statistics (mean velocity and TI), with turbine power calculated from the predicted wake velocity field. A high-fidelity dataset is generated via large-eddy simulation (LES) for a full-scale 120 kW TCT under nine inflow conditions, combining three velocities and three TI levels, forming a condition-disjoint benchmark for generalization. To ensure physical consistency, the model integrates LES-inspired plane-wise weak regularization, including boundary-consistency constraints, momentum-balance residuals, WALE-inspired dissipation, and TI-consistent TKE-budget constraints. Training robustness is enhanced through warmup, sparse physics-point injection, dynamic loss weighting, and Bayesian hyperparameter optimization. Under a strict leave-one-condition-out protocol, the proposed framework achieves mean R2 of 0.9953 for wake velocity and 0.9774 for wake TI, while power predictions maintain <1.2% relative error across all folds. These results demonstrate a scalable and physically consistent surrogate for rapid wake–power evaluation and practical array-level optimization under heterogeneous inflow environments.