Where Price-Aware Wake Steering Sends a Neural Surrogate

Wind farms can raise total output by turning upstream turbines slightly away from the wind, a practice called yaw (wake) steering, so that their slow, turbulent wakes miss the turbines behind them. Choosing those angles requires a flow model fast enough to call thousands of times, which is why fast neural networks that imitate a slower simulator are now used in its place. These neural surrogates, together with the optimizers that call them, have been built under two silent assumptions. The first is that every megawatt-hour is worth the same: a static price. The second is that the wake is essentially a two-dimensional slice at hub height. This note argues that the two assumptions fail together. Once the optimizer weighs a time-varying electricity price against the mechanical wear that yawing causes, its preferred angles move with the price. During high-price intervals it pushes toward aggressive yaw. That is exactly the region where three-dimensional effects dominate, such as the curled wake, the asymmetry caused by wind shear, and blade stall on one side of the rotor. It is also the region the surrogate's training data covers worst.

The core claim is that, for a surrogate used inside an operating decision loop, the out-of-distribution problem that matters is set by the price process, not by the layout or the weather alone. We propose to measure this shift and show that ordinary field-error metrics hide it. We then test whether a simple fix closes the gap: spend the expensive three-dimensional simulation budget along the path of price-dependent optimal decisions.

Why price changes the decision, and when it does not

A reviewer's first objection deserves an answer up front. If revenue is just price times power, price multiplies the whole objective and the optimal yaw angles do not depend on it, apart from curtailing when prices are negative. Dynamic pricing only matters when another term enters the objective in different units. Wear is such a term. Let $\gamma$ be the vector of yaw offsets, $s_t$ the wind state in interval $t$ (direction, speed, turbulence intensity, shear exponent), $p_t$ the price, $P_i$ the power of turbine $i$, $D_i$ a damage-rate proxy for turbine $i$'s yaw drive and rotor, and $c$ the monetary cost per unit damage. The per-interval objective is

$$J_t(\gamma) = p_t \sum_i P_i(\gamma, s_t) - c \sum_i D_i(\gamma, s_t).$$

For $p_t > 0$, the maximizer depends on the price only through the ratio $\kappa_t = c/p_t$. Price-aware steering is therefore a one-parameter family of yaw policies indexed by $\kappa$. Low $\kappa$, meaning high prices, calls for aggressive steering. High $\kappa$ calls for little or none. The literature on active control and fatigue motivates this framing: Barradas Berglind et al. and Gong et al. report that aggressive actuation shifts load onto actuators. Monroc shows that steering optimizers push toward large yaw angles when wear is free. Because the true form of $D_i$ is uncertain, the program treats it as a sensitivity axis rather than a known law.

Relation to prior work

Three surrogate papers supply the method we build on and the gap we target. Li et al. train a graph transformer on analytical PyWake wakes and generalize across layouts. WakeNet pre-trains on Gaussian-wake data and fine-tunes on 100 Curl-model samples, and its authors name the extension to three-dimensional fields with vertical shear as future work. Kayedpour et al. is the closest work and the one we must clearly differ from. They already couple market bidding with wake control, using a fuzzy-neural surrogate of FLORIS's Gauss-Curl-Hybrid model on one layout. They report up to 12% efficiency gain over uniform reserve allocation. Their surrogate is fit on a fixed precomputed grid, and they do not ask whether the market-driven optimizer queries it outside that grid. Our proposal does not add another coupled market controller. It studies the surrogate as the weak link: we measure how the price process reshapes the surrogate's query distribution and what that does to the money actually lost.

Two lines of work constrain the design. AirfRANS shows that graph surrogates track lift well but lose drag, with poor or negative rank correlation, and that errors spike outside the training Reynolds number and angle-of-attack range. This warns that good aggregate field error can coexist with failures in derived engineering quantities. Beucler et al. show that enforcing conservation to machine precision is necessary but not sufficient for extrapolation. That argues against expecting physics constraints alone to cover the aggressive-yaw regime.

Fluid mechanics motivates why the shifted region is physically different, not merely sparse. Kamada et al. measure an "azimuthal inequality" on a yawed rotor, with stall on the downstream sector at low tip-speed ratio. Their power curves are symmetric in $\pm$yaw, but that was in a tunnel without atmospheric shear. Real shear should break this symmetry, and a hub-height slice model cannot represent the asymmetry. On the methods side, the Buckingham π test-time projection rescales out-of-range inputs onto matching dimensionless training cases and cut extreme out-of-distribution error by up to 91% on 2D heat conduction and elasticity. We reuse it, and we expect it to help along wind speed but not along shear or turbulence intensity. Those variables are themselves dimensionless, so rescaling cannot move them back into the training range.

Research Program

Ground truth and three-dimensional effects. We set up a fidelity ladder honestly. The reference model is FLORIS Gauss-Curl-Hybrid with vertical shear and veer enabled. It contains the curled-wake and shear physics that hub-height Gaussian models lack. It is cheap enough to compute exact regret over a full year of 15-minute intervals, yet expensive enough that per-interval optimization at farm scale is slow (Li et al. report 45 minutes for an 80-turbine genetic-algorithm optimization with analytical models). The low-fidelity source is the Gaussian wake model. A small set of large-eddy simulations (LES), which resolve the large turbulent eddies directly and serve as the closest available stand-in for real flow, at selected aggressive-yaw states acts as a sanity check that the reference model's high-yaw behaviour is not itself the artefact. Fatigue enters as a damage proxy that grows with yaw magnitude and with sheared-inflow load cycles. We test three proxy shapes.

Measuring query shift. We take a WakeNet-style surrogate trained with the standard recipe: fine-tuning samples drawn uniformly or around maximum-power yaw schedules. We then build the $\kappa$-indexed yaw lookup table through it. For every interval of a held-out price year, we record where the optimizer queries the surrogate relative to the training support, and the regret in net value (revenue minus wear cost), evaluated on the reference model. The analysis tests whether regret concentrates in low-$\kappa$ intervals and in sheared or high-turbulence states, and whether field error predicts regret.

The proposed fix. Decision-path acquisition proceeds iteratively. We solve for optimal yaw across a grid of $\kappa$ values with the current surrogate. We then spend the next batch of expensive simulations at the states and angles those solutions actually visit, and refit. Optimization is also restricted to stay near the training support, with a distance penalty on the support, so that the optimizer cannot exploit regions where the surrogate is unchecked. Buckingham π projection is applied along wind speed as a separate component.

Baselines

The naive reference point is greedy operation: zero yaw offset everywhere, with curtailment when the price is negative. Every method must beat it in net value, or steering is not worth its wear.

The static-price baseline is the current practice embodied by Li et al. and WakeNet. It uses a maximum-power yaw table, with price used only for curtailment. It isolates the cost of ignoring $\kappa$.

The strongest surrogate competitor is WakeNet's multi-fidelity transfer recipe, retrained by us on the same Gaussian pre-training set and given exactly the same number of reference-model fine-tuning samples as our method. It receives the same architecture, the same hyperparameter search budget, and uniform or maximum-power sampling in place of decision-path sampling. For context on the accuracy level these methods reach, Li et al. report 99.8% average relative power accuracy on unseen layouts, and WakeNet reports 99.8% accuracy against FLORIS. No reported result was found on Lacuna for net-value regret of any wake surrogate under dynamic prices. Beating this baseline means at least halving net-value regret on low-$\kappa$ intervals at equal simulation budget, while matching it on high-$\kappa$ intervals.

The oracle is the $\kappa$-indexed table computed directly with the reference model offline. It defines zero regret.

The key ablations remove decision-path acquisition (keeping the support penalty), remove the support penalty, and remove π projection. Each ablation answers one question about which ingredient carries the gain.

Why not just…?

Why not just use FLORIS directly to build the price-indexed lookup table offline? This is a real threat. When the reference model is FLORIS, direct offline tabulation over wind state and $\kappa$ is the oracle, and it may be affordable. The proposal survives only to the extent that the true reference is costlier than FLORIS. LES is far too expensive to tabulate, and the extra $\kappa$ axis multiplies the grid size. We will report wall-clock cost for direct tabulation against surrogate tabulation. If direct FLORIS tabulation is cheap enough at Horns Rev 1 scale, we will say so. The contribution then narrows to the measured query shift and to acquisition under an LES budget.

Why not just train on more uniformly sampled data or use an ensemble for uncertainty? Uniform sampling spends most of its budget in yaw regions no optimizer visits at the prevailing prices. Ensembles flag disagreement but do not, by themselves, add data where decisions land. The equal-budget comparison with WakeNet-style uniform sampling, and an ensemble variant of that baseline given the same budget, settles this directly.

Why not optimize power first and pass it to a dispatch layer? With wear in the objective, the optimal yaw depends on $\kappa$, so a power-first table is optimal only at $\kappa = 0$. Kayedpour et al. already found that coupling beats a decoupled reserve allocation. Our static-price baseline measures the size of this loss in net value.

Why not enforce physical constraints instead? Beucler et al. show that exact conservation does not prevent degradation under regime shift. AirfRANS shows that physics-aware graph models still fail on wall-gradient quantities when extrapolating. We therefore do not expect constraints to supply missing aggressive-yaw data, though they could be combined with our method.

Proposed Experiments

The testbed is the 80-turbine Horns Rev 1 layout used by Li et al., with wind states sampled across all directions. Shear exponent and turbulence intensity are held out at the high end to create physical out-of-distribution test sets. Prices come from the ENTSO-E series aggregated in Real-E, using the Danish zones. Real-E documents that these series are volatile, with temporal graph volatility of 1.03 against 0.58 for older datasets. We train on one year and test on a later one. If Real-E lacks day-ahead prices for these zones, a separate ENTSO-E price extraction is a dependency we flag now.

The metrics are net-value regret relative to the oracle, broken down by $\kappa$ decile and by in- versus out-of-distribution wind state. Alongside regret we report field and power error, and the rank correlation between field error and regret across all trained surrogates. A small-correlation finding would mean that the evaluation practice of Li et al. and WakeNet misses what operators care about.

Three decisive checks would make the case. First, static-sampled surrogates show regret concentrated at low $\kappa$ and high shear. Second, decision-path acquisition removes most of that regret at equal budget, and the support penalty explains the remainder. Third, the LES spot checks agree with the reference model on the sign of the shear-induced yaw asymmetry, which confirms that the regime being chased is physical. If none of these hold, and uniform sampling matches our method, we will report that the static-pricing assumption is harmless for surrogate design.