Predicting wind speeds in complex forested terrain using Reynolds-Averaged Navier-Stokes surrogates and data-driven feedforward convolutional neural networks

Predicting onshore wind farm performance in complex terrain remains difficult due to variable wind behavior, forest heterogeneity, and topographic complexity. Accurate site assessments typically require numerous microscale Computational Fluid Dynamics (CFD) simulations across a wide parameter space, but these simulations are computationally expensive and impractical for exhaustive sweeps. This study investigates whether data-driven surrogates can leverage CFD outputs to provide accurate, efficient wind speed predictions in such environments.
We train Reynolds-Averaged Navier–Stokes (RANS) surrogate models using data-driven neural networks using CFD-RANS wind fields over forested terrain, generated using the OpenFOAM-based FIWIND toolchain for a range of forest heights and wind directions. Model inputs consist of forest height–wind direction–wind speed profile tuples at selected terrain locations, with wind speed as the output. A feedforward convolutional neural network is trained with mean squared error loss, incorporating residuals from CFD results, and optimized using L-BFGS and Adam algorithms. Surrogate predictions are benchmarked against CFD simulations produced with the OpenFOAM-based FIWIND toolchain.
The surrogates are constructed using different subsets of training data derived from vertical and horizontal wind speed profiles. Results demonstrate strong extrapolation capability: vertical wind speed profiles are reproduced within 10% error under unseen conditions, and horizontal wind fields at hub height are predicted within 15% relative error at a 99% confidence level. While performance could be further improved through physics-based constraints or more advanced architectures, our findings indicate that purely data-driven RANS-based surrogates are straightforward to construct and offer promising generalization capabilities. To enhance robustness, we extended the training dataset to incorporate two spatial resolutions, making the approach particularly well suited as a foundation for multi-fidelity surrogates in domains with varying governing physics.