Dispatching Wind Farms Around Their Own Resonances

Wind farms are increasingly asked to do more than produce as much power as possible. Grid operators want them to follow a power reference that changes minute to minute and to hold back reserve power for frequency emergencies. Each such instruction changes how fast each rotor spins. Every rotor sends a periodic pulse into its tower at three times its rotation frequency, once per blade pass, called the "3P" frequency. When that pulse matches a natural frequency of the tower or of a floating platform, the structure resonates. A dispatch that looks harmless on a power-balance spreadsheet can then leave turbines vibrating at resonance for hours. The cost of a set-point is therefore a sharply bumpy function of rotor speed: flat almost everywhere, with narrow, tall peaks. Convex dispatch models replace those peaks with a smooth bowl, and their solutions can sit right on top of a peak.

We propose to treat this failure of convexity as a specific, exploitable structure rather than as generic "non-convexity" to be attacked with swarm search or learned smooth surrogates. The cost is non-convex in only one way: each turbine must choose which side of each resonance band to operate on, and inside each choice the problem is an ordinary convex quadratic program. Wakes, the slowed and turbulent air behind upstream rotors, couple these choices only between neighbouring turbines. We build a dispatch method that makes the band choices exactly and solves the rest convexly. We then measure what convex and heuristic dispatchers actually lose on simulated farms driven by real intermittent wind. The headline deliverable is that measurement, the "price of convexity" in tower fatigue and reserve delivery. If the price is small, that negative result is useful too.

Core Claim

Over-speed deloading is a standard way to store kinetic energy for Frequency Containment Reserve (FCR), the fast reserve that stabilises grid frequency in the first seconds after a fault. An Optimal Wind Farm Operation Strategy for the Provision of Frequency Containment Reserve Incorporating Active Wake Control uses it, with particle swarm search over wake-aware deloading levels. Over-speed deloading raises rotor speed by design. On "soft-stiff" towers, whose first natural frequency sits between the 1P and 3P excitations, this can drive 3P straight into the tower mode. FLOAT shows how serious that coupling is under normal production. The baseline IEA 22 MW tower on a floating platform reaches its fatigue limit in about 9 months. Moving it to a stiff-stiff configuration that avoids 3P restores a 25-year life, at a 68.8% mass increase.

Wakes add a second excitation path. In LES (large-eddy simulation, a high-resolution turbulent flow solver) of two 5.5 MW turbines in tandem, Numerical Simulation of Large Scale Wind Turbine Wake and the Influence on Vibration of Downstream Wind Turbine finds the downstream turbine's vibration spectrum peaking at 0.4131 Hz, against a 3P frequency of 0.415 Hz. Field data from the same work give R² = 0.728 between axial vibration frequency and 3P. The resonance risk of a downstream turbine therefore depends partly on its upstream neighbour's set-point.

Our claim is that farm-level dispatch should encode these resonance bands as disjunctive constraints, meaning constraints that are satisfied by being on one side or the other. Doing so reduces tower fatigue during reserve provision and power tracking at little or no cost in tracking accuracy or reserve. Both linearised convex model predictive control (MPC) and smooth non-convex surrogates should miss most of that gain.

The Formulation

For turbine $i$ with tower or platform modal frequencies $f_{i,k}$, the resonance band in rotor speed $\Omega$ is

$$ B_{i,k} = \{\Omega : |3\Omega/2\pi - f_{i,k}| < \delta_{i,k}(\theta_{\text{up}(i)})\}, $$

where $\delta_{i,k}$ is the band half-width. The width depends on the set-points $\theta_{\text{up}(i)}$ of turbine $i$'s upstream wake neighbours, because their periodic wake loading strengthens the excitation. We identify $f_{i,k}$ and $\delta_{i,k}$ from aeroelastic simulation; the method does not assume them.

Dispatch over a horizon then minimises a convex tracking-plus-fatigue quadratic. It also adds a dwell penalty $\sum_t \mathbf{1}[\Omega_{i,t} \in B_{i,k}]$. The penalty is needed because sweeping quickly through a band during a reserve ramp is cheap, while parking inside one is expensive. One binary variable per turbine, band and decision block encodes which side of the band the turbine sits on.

Two facts keep this tractable. First, each turbine has only one or two relevant bands. Second, band widths couple only along the sparse wake graph that Scalable Optimization for Wind Farm Control using Coordination Graphs already shows can be built from FLORIS geometry. FLORIS is a steady-state engineering wake model.

We solve the problem as a mixed-integer quadratic program. To decompose by wake neighbourhood, we relax the farm-wide power constraint with a single price variable, so each neighbourhood can be solved separately for a given price. Inside each fixed choice of sides, the problem stays convex, so solutions are certified rather than heuristic.

Research Program

The work has three parts. The first is band identification. We run OpenFAST on the IEA 22 MW reference turbine and the NREL 5 MW turbine across sweeps of deloading level, deloading mode (over-speed versus pitch) and upstream wake inflow. From these runs we extract $f_{i,k}$, $\delta_{i,k}$ and tower damage-equivalent loads, using the Monte Carlo load-case sampling of FLOAT. A small learned regressor maps upstream thrust and spacing to band width. This is the only learned component, and we report its error separately.

The second part is the dispatcher above, run in receding horizon. That means re-solving every control interval and applying only the first step.

The third part is evaluation in SimWindFarm on the 10-turbine and 100-turbine Thanet layouts used by Model Predictive Controllers for Reduction of Mechanical Fatigue in Wind Farms, and on the C-Power layout used for FCR. Intermittency comes from wind speed and direction traces in the SDWPF dataset, 134 turbines over 245 days at 10-minute resolution, upsampled with Kaimal-spectrum turbulence. A sample of dispatch decisions is re-simulated in OpenFAST, so that fatigue claims do not rest on the dispatcher's own model.

The tension between two prior studies matters here. Modal Analysis of the Wake Shed Behind a Horizontal Axis Wind Turbine with Flexible Blades finds far-wake modal energy within 2% between flexible and rigid blades, and argues that rigid-blade wake models suffice. The tandem LES study finds strong 3P coupling downstream. Our design handles both possibilities. Band widths can depend on upstream set-points, but the ablation that removes this dependence tests directly whether wake-coupled resonance matters at realistic spacings. If the rigid-wake view is right, the method simplifies to independent per-turbine bands, and we will report that.

Baselines

The naive reference is the open-loop proportional scheduler in SimWindFarm, which splits the farm reference by available power. This is the reference that Model Predictive Controllers for Reduction of Mechanical Fatigue in Wind Farms improves on. That paper reports 22.30% lower main-shaft fatigue and 4.75% lower tower bending moment on 10 turbines, and 13.07% and 2.13% on the 100-turbine Thanet farm.

Its deterministic linearised MPC is the strongest convex competitor. We run it with the same ARMA turbulence prediction (autoregressive moving average, a classical short-term time-series predictor), the same horizon and the same compute. We tune its fatigue weights with the same budget we use for our method. Beating it means lower tower damage-equivalent load, verified in OpenFAST, at equal or better power-tracking RMSE. Because the reported tower reduction is only 2–5%, we expect most of our margin on the tower channel. A gain under roughly 1% of that reduction would count as not beating it.

For reserve provision, the strongest existing method is the wake-aware ANFIS-surrogate plus particle swarm strategy of An Optimal Wind Farm Operation Strategy for the Provision of Frequency Containment Reserve Incorporating Active Wake Control. ANFIS is an adaptive neuro-fuzzy inference system, a small neural surrogate. That paper reports up to 12% better reserve delivery efficiency than uniform deloading. We re-run it on the C-Power layout with an equal wall-clock budget. Beating it means matching its reserve efficiency within 1 percentage point while reducing resonance dwell time and tower damage. We found no reported tower-fatigue number for it on Lacuna, so we measure that ourselves.

We also include Set-Point Thompson Sampling from Scalable Optimization for Wind Farm Control using Coordination Graphs, given our band labels as its discrete regimes. It reports a 100% success rate in avoiding damage-inducing set-points, so on avoidance alone it may tie us. We expect it to lose on tracking accuracy because its set-points are discrete.

Finally, we include a smooth learned fatigue surrogate in the style of End-to-end Stochastic Optimization with Energy-based Model, which reports 7.3% lower task loss than predict-then-optimize on stylised power scheduling. We train it on the same OpenFAST data.

Three ablations isolate the key ingredients. One replaces the disjunctive bands with their convex hull. One removes the wake dependence of band width. One drops the dwell term and keeps only static band avoidance.

Why not just…?

Why not just put rotor-speed exclusion zones in each turbine's local controller and keep farm dispatch convex? This is the most serious threat, because exclusion zones are standard practice. The weakness is that a local controller can only refuse a set-point. It cannot pass that power to a neighbour. The result is either a tracking error or a turbine pinned at the band edge with a different wake than the farm model expected. We implement this baseline exactly: convex farm MPC plus local exclusion with fast band crossing. If it comes within our confidence intervals on fatigue, tracking and reserve, we conclude that the convexity assumption is harmless at farm level, and we report that as the finding.

Why not just linearise and re-linearise, solving a sequence of convex programs? Repeated linearisation around the current point stays on whichever side of a band it starts on. It cannot jump across one, because the gradient points toward the nearer edge. Model Predictive Controllers for Reduction of Mechanical Fatigue in Wind Farms already notes that its linearised controller degrades beyond 2–3 prediction steps. The convex-hull ablation measures this loss directly.

Why not just use particle swarm search or Bayesian optimisation over the full non-convex problem? These methods give no bound on solve time, which matters when an FCR event demands a response within seconds. They also give no certificate of how far a solution is from the best one. We compare against the swarm method at matched wall-clock time. We report the optimality gap our solver certifies on the same instances, and the distribution of solve times on 100 turbines.

Why not just learn a smooth non-convex surrogate, energy-based or gradient-norm style? Smooth surrogates blur narrow resonance peaks. Gradient-norm methods such as DINGO guarantee convergence only on invex landscapes, where every stationary point is a global minimum. A banded cost is not invex, because the flat regions between peaks are stationary everywhere. The surrogate baseline tests this. We predict it will show more dwell time inside bands than the disjunctive method.

Decisive Checks

The first check is the price of convexity. On Thanet driven by SDWPF wind, we measure the fraction of turbine-hours that convex MPC parks inside identified resonance bands. We also measure the tower damage-equivalent load difference between convex MPC and our method, verified in OpenFAST for both the 5 MW and 22 MW soft-stiff configurations. A null result here ends the claim.

The second check is the timing deadline. The disjunctive solve on 100 turbines must finish within the FCR control interval at the 95th percentile. We report the observed distribution because no prior number exists.

The third check is the wake-coupling ablation. It shows whether the coupling observed in tandem LES changes dispatch decisions at realistic spacings, which settles the disagreement between the two wake studies for operational purposes.

The fourth check is robustness. We perturb identified band centres by ±1–3% and widths by ±20%. The comparison is whether our method's advantage over the local exclusion-zone baseline survives the kind of model error that a learned band regressor will have.