For AI models predicting wind turbine stress, feeding them a “richer picture” of the wind matters more than making the model smarter
Predicting how much power an offshore wind farm produces — and how much stress its turbines are accumulating — ideally happens fast enough to inform real-time control and long-term design decisions. High-fidelity aeroelastic simulations can do both accurately, but at a computational cost that makes them impractical at scale. Machine-learning surrogates offer a faster alternative, and researchers have spent years refining their architectures to close the accuracy gap.
A new study suggests the architecture may not be the most important lever to pull. How wind data is represented before it reaches the model turns out to matter enormously, and not in the same way for every quantity being predicted.
Why wind farms need faster predictions
High-fidelity aeroelastic simulations are the gold standard for estimating turbine power and structural fatigue, but running them at scale is computationally prohibitive. Wind farm optimization, real-time control, and digital-twin applications all require predictions repeated across hundreds or thousands of operating scenarios. That cost barrier is exactly why machine-learning surrogate models have attracted growing interest: once trained, they approximate aeroelastic outputs at a fraction of the computational expense.
The baseline, full-grid, and PCA configurations were tested using the same simple neural network, so any performance differences reflect the input — not the model.
This study uses the publicly available HAWC2Farm Lillgrund dataset — 48 offshore turbines across 800 simulated scenarios, covering a wide range of wind speeds, turbulence intensities, and shear conditions. Rather than proposing a new architecture, the researchers focused on a less-examined question: does the form of wind data fed into the model matter as much as the model itself?
Four ways to describe the wind hitting a rotor
The rotor plane — the disc swept by a turbine’s blades — can be described with very different levels of spatial detail. The study tests four distinct approaches.
The simplest, called the baseline, reduces everything to two numbers: average wind speed and average wind variance across the rotor. At the other extreme, the full-grid representation retains all 1,800 measurements from a 30×30 sampling grid. Between those poles sit two intermediate options: Principal Component Analysis (PCA) compresses the full grid to 128 key variables capturing dominant variation patterns, while a convolutional neural network (CNN) receives data as a 2D spatial image, preserving neighborhood relationships between adjacent grid points.
The baseline, full-grid, and PCA configurations were tested using the same simple neural network, so any performance differences reflect the input — not the model.
Power prediction: two numbers are almost enough
For electrical power, the two-variable baseline was surprisingly competitive, achieving an RMSE of roughly 2.2% of rated power. Richer inputs reduced that error further, though the gap was modest.
PCA compression to 128 components delivered the best result among fixed-architecture comparisons, cutting RMSE to approximately 0.24% of rated power — outperforming the full 1,800-variable input under the same network. Stripping out redundant variation apparently helps the model focus on what matters. A deeper residual MLP actually performed worse than the simpler model on identical data, suggesting that adding model capacity doesn’t compensate for an unhelpful input representation. The topology-aware CNN also didn’t achieve the lowest power error. For this task, a compact statistical summary of the rotor plane is sufficient.
Fatigue loads: spatial detail can make or break the prediction
Structural fatigue tells a very different story. The benefit of richer wind-field data varied dramatically across the 12 load components studied.
For tower-base fore-aft bending, switching from the two-variable baseline to the full grid dropped RMSE only modestly — from 117 to 111 kNm. That channel is dominated by rotor-integrated thrust, which rotor-averaged quantities already capture reasonably well. Blade-root edgewise bending and main-bearing bending behaved differently: errors fell by factors of roughly five and six respectively when spatial wind-field data replaced the baseline. Those components are sensitive to asymmetric rotor loading — exactly the structure that distributed inflow information captures and a two-number summary cannot.
Across all 12 fatigue-load components, the topology-aware CNN achieved the lowest prediction errors, confirming that spatial architecture matters most when the target is structural rather than electrical.
One model does not fit all prediction tasks
No single input representation or model architecture was optimal for both prediction tasks. PCA-compressed input paired with a simple multilayer perceptron worked best for power; the spatial CNN worked best for structural loads. Using the same setup for both means accepting a performance penalty on at least one.
The study also has honest limitations. It relies on a single simulated dataset, and the wind fields are temporally aggregated rather than time-resolved. Whether these findings hold across different turbine designs, farm layouts, or real field measurements remains open. Transferring simulation-trained surrogates to operational settings would mean dealing with measurement noise, spatial sparsity, and other complications the Lillgrund dataset doesn’t include.
What the study establishes is a clearer design principle: choose your input representation based on what you’re trying to predict. As surrogate models move closer to deployment in optimization loops and digital twins, matching the data format to the prediction target may prove just as consequential as any advance in model architecture.
The complete study is available here: Ganapathi Subramanian, P.: The influence of flow-field representation on electrical power and fatigue-load surrogate modelling in offshore wind farms, Wind Energ. Sci., 11, 4041–4057, https://doi.org/10.5194/wes-11-4041-2026, 2026.
Daniel García is an Editor-in-Chief with strong expertise in structural work and engineering principles. He combines this technical foundation with deep knowledge of energy, spatial design, and emerging technologies, bringing a forward-thinking and analytical approach to editorial leadership.
