Region-aware hybrid LSTM-GRU modeling for wake flows induced by offshore wind-turbine foundations

• A region-aware reduced-order modeling framework is proposed by integrating Grad-CAM–based wake localization with POD and hybrid LSTM–GRU temporal prediction. • Automatic wake-region detection effectively isolates dynamically significant flow structures, enhancing spatial focus and reducing redundant far-field content. • Dual POD decomposition and a fused LSTM–GRU architecture capture multi-scale wake dynamics, improving temporal coherence and vortex-shedding prediction. • The fused wake–global reconstruction reduces the relative reconstruction error from 24.25% to 15.61%, outperforming global-only ROM approaches. The wake dynamics downstream of offshore wind turbine foundations exhibit strong nonlinearity, temporal variability, and three-dimensional disturbances. Accurately and efficiently modeling these unsteady flows has long been a challenge. Traditional recurrent neural network (RNN) methods often struggle to capture localized, nonstationary wake structures due to limited spatial awareness and sensitivity to variations in input dimensionality. This study presents a region-aware hybrid modeling framework to address these limitations. The approach integrates three key components: automatic wake-region detection, Proper Orthogonal Decomposition (POD) for dimensionality reduction, and a hybrid LSTM-GRU network for temporal sequence modeling. A CNN enhanced with Grad-CAM identifies dynamically active wake regions. POD is applied separately to both the global domain and the extracted wake subdomain to derive dominant modal coefficients. A dual-branch temporal prediction model then forecasts future low-dimensional flow representations for both domains. The training process employs a composite loss function combining time-domain mean squared error (MSE) with frequency-domain Fourier loss. To prioritize key unsteady features, higher weighting is applied to wake-region predictions. Results show that the proposed model effectively captures key wake dynamics, including vortex shedding patterns, streamwise wake decay, and localized disturbances. For full-field predictions, the mean absolute relative error (MARE) is reduced from 15.6% for the conventional global POD with a standard LSTM baseline to 12.2% for the final blended prediction of the proposed region-aware hybrid framework, indicating a substantial improvement in predictive accuracy. The applicability of the proposed approach to different Reynolds numbers, alternative foundation geometries, and unsteady tidal inflow conditions remains to be investigated.