A Hybrid physics-informed digital twin for offshore wind-farm flow prediction: integrating CFD simulations with deep regression forest and probabilistic surrogate models

The marine-atmosphere dynamics are nonlinear and multiscale, which poses a computational challenge for accurately describing the flow field around offshore wind turbines. Using advanced data-driven surrogate models and CFD simulations, this study presents a hybrid physics-informed digital twin framework to predict wind farm behavior under stochastic environmental conditions with high accuracy. In this work, we propose a framework that simultaneously integrates spatiotemporal dependencies, physical constraints, and probabilistic uncertainties by combining Physics-Informed Neural Networks (PINNs), Deep-Gaussian Process Regression (Deep-GPR), Extreme Gradient Boosting Algorithm (XGBoost), and a transformer-based regression model called CFD-Former. PINNs incorporate Navier–Stokes equation residuals into the training process to ensure physical consistency, while XGBoost and Deep-GPR improve model generalizability to different inlet flow conditions. Hybrid surrogate models are trained on data from CFD calculations with different wind direction and speed conditions. The generated digital twin is able to predict the velocities measured at the meteorological tower within an uncertainty of 10% and reproduce the high resolution flow patterns due to CFD with a mean absolute deviation below 1%. This novel approach uses deterministic simulation and intelligent inference to predict offshore wind farm performance in a scalable and physically based manner, possibly enabling adaptive operation and reliability assessment.