Hybrid AI-reduced-order model for real-time wind flow prediction around offshore wind farms

Accurate characterization of wind flows in harsh environments such as offshore wind farms is crucial to ensure the safety and efficiency of Unmanned Aerial Vehicles (UAVs) operations in inspection and maintenance tasks. However, Computational Fluid Dynamics (CFD) simulations remain prohibitively expensive for real-time applications. This work presents a surrogate modeling framework that combines Convolutional Autoencoders (CAE) with a Deep Neural Network (DNN) to enable fast prediction of complex wind fields and assess their influence on UAV trajectory planning. A database of Reynolds-Averaged Navier–Stokes (RANS) simulations around an offshore wind turbine, parametri-zed by boundary-condition descriptors, serves as the training foundation. Two-dimensional horizontal slices at different heights are extracted from each simulation and used as input fields for the CAE, which compresses the velocity magnitude and turbulent kinetic energy distributions into a low-dimensional latent space. The DNN then maps the inflow speed, direction, and slice height to this space, enabling 3D field reconstruction by combining predictions at multiple vertical levels. The flow reconstructions obtained through the predictive model reproduce key flow features with mean absolute errors around 1 × 1 0 − 3 in most regions, with higher deviations confined near walls. The surrogate achieves a computational speed-up of approximately 1 . 9 × 1 0 5 compared to RANS-CFD, enabling near real-time inference on embedded hardware. The predicted fields are integrated into a path-planning framework to identify safe and efficient trajectories by evaluating local turbulence and wind exposure. By merging the surrogate model with flow-aware navigation, the study bridges the gap between the accuracy of the CFD-level and operational scalability in UAV-based offshore maintenance.