Modeling GW-Scale Offshore Wind Arrays: What Changhua 2b & 4 Teaches Us About Array Efficiency

Ørsted has commissioned the Greater Changhua 2b and 4 offshore wind farms in the Taiwan Strait — combined, on the order of 1 GW of new offshore wind nameplate delivered roughly 35–50 km off Taiwan’s Changhua coast. Over the full Greater Changhua build-out — phases 1, 2a, and now 2b & 4 — the project zone totals more than 2 GW, making it one of the largest operational offshore wind clusters in Asia Pacific (Ørsted project communications).

The Resource: Why the Strait Changes the Modeling Baseline

Before any wake or cable model matters, the wind resource sets every downstream number. The Taiwan Strait is a channel between the Asian mainland and Taiwan’s mountainous spine, and that topography funnels the seasonal monsoon and the northeast trade wind into a reliable, high-energy corridor. Long-term offshore measurements in the Changhua window consistently return annual mean wind speeds at hub height in the 10–11 m/s class — comparable to, and in places better than, the North Sea sites that anchored early European offshore growth, and far above the 8–9 m/s typical of many U.S. East Coast and Japanese lease areas.

Two modeling consequences follow directly from that resource:

• Higher energy density per installed MW. At an 11 m/s mean, the cube-law in the power curve turns marginal wind-speed differences into materially different gross yields. A resource error of just 0.5 m/s at this level moves gross annual energy by roughly 10–12%, which is larger than most individual loss terms a model tries to capture — so the resource input dominates every downstream uncertainty.
• Wake recovery behaves differently. Strong ambient turbulence and the strait’s thermal stratification mean wakes mix and recover faster than they would in weakly turbulent, stably stratified North Sea summer conditions. A model tuned to one site and applied to another will misstate wake loss — this is the single most common error I see in transferred offshore models. Wake-loss is not a universal constant; it is a function of the local turbulence regime, and the Strait’s regime sits at the favorable end of the spectrum.

It is precisely because the gross resource is so good that array-level losses matter. A 6–8% wake loss on a low-resource site is annoying; on a Changhua-scale site running at a 48% gross capacity factor, that same percentage is hundreds of GWh per year left on the table.

Deep-Array Wake Loss: The Scaling That Breaks Simple Spreadsheets

Single-row and single-turbine wake thinking is fine for a 20–40 MW onshore wind farm with one or two turbines per row facing the dominant wind. A GW-scale offshore array is a different object. Changhua 2b & 4 alone deploy on the order of one hundred 9–11 MW class turbines (typical of the recent Taiwan Strait build-out), laid out in blocks where interior machines sit many rows behind a windward edge. In those interior rows, wake loss does not simply add — it compounds, because every upwind turbine has itself already been slowed by its own upwind neighbors. This “deep-array effect” is why measured aggregate wake losses at large offshore arrays routinely land in the 6–10% of gross-annual range even at well-optimized 7–8 rotor-diameter (D) spacings, rather than the 2–4% a single-wake superposition would predict.

Downwind row spacing is the strongest single layout lever. Figure 1 illustrates the sensitivity of modeled interior-array wake loss to starting-row spacing for a deep-layout block of a Changhua-scale array. These are my modeling assumptions for a typical turbulence regime in the Strait — not Ørsted’s figures — but they represent the standard shape of the curve: losses rise steeply as spacing tightens toward 5–6 D, where full-wake alignment across many rows dominates, and flatten as spacing opens toward 10 D:

The catch is that tightening crosswind (along-row) spacing to fit more capacity onto a constrained lease area pulls array wake losses back up even when downwind spacing looks generous — so the true optimization is two-dimensional and interaction-driven. That is precisely the kind of problem a flat spreadsheet multiplication of “3% wake” cannot represent, because in a real deep array there is no single wake-loss number; there is a distribution of wake losses across every turbine in every row, varying hour by hour with wind speed, direction, and atmospheric stability.

The Full AEP Budget at Changhua Scale

Moving from gross to net annual energy at a project of this scale means stacking a consistent loss budget. Figure 2 shows an illustrative full-array AEP budget for a ~1.1 GW, ~100-turbine Changhua-scale array — again my modeling baseline, spelled out so the arithmetic is auditable rather than buried:

• Gross output (index = 100): the theoretical sum of every turbine’s power curve applied to the local met mast data, before any interaction loss.
• − Array wake (~6 pts): aggregate interior wake loss for a 7–8 D downwind, ~3–4 D crosswind layout tuned to strong-turbulence recovery. This 6% is net of the “deep-array saturation” discussed above.
• − Availability (~3 pts): offshore turbine and balance-of-plant downtime in the Taiwan operational context, where typhoon-driven planned downtime and marine access windows are material.
• − Electrical losses (~2.5 pts): array cable I²R losses, transformer and converter station losses across the collector and export path.
• − Curtailment & other (~1.5 pts): grid-side curtailment, blade soiling, and turbine control offsets between the ideal power curve and field operation.

Net of those four terms lands at an index of ~87, or roughly a 13% array-level loss stack on top of a gross figure that is itself already excellent:

The key modeling discipline figure 2 is built to illustrate: each loss term has a different sensitivity and different uncertainty, and they interact. Tightening spacing saves electrical cable length but raises wake loss. Pushing the export voltage higher cuts electrical losses but adds transformer cost and complexity. A credible model has to let all of these co-evolve rather than fixing each at a constant percentage.

Array Collector Sizing and Offshore Export Architecture

At GW scale the electrical network stops being an afterthought and becomes a design variable with real yield and CAPEX consequences. Two coupled decisions dominate:

Collector (array) network voltage and layout

Offshore projects at this scale typically group turbines onto medium-voltage collector strings — conventionally 33 kV, increasingly 66 kV on modern offshore projects. The arithmetic is straightforward: doubling collector voltage cuts string current for the same power by half, which cuts I²R losses by roughly a factor of four for the same cable cross-section, or lets the designer use thinner, cheaper cables for the same loss. The modeling job is not to pick “33 or 66” in the abstract, but to optimize string length, number of turbines per string, cable cross-section under current-carrying (ampacity) limits, and the routing that minimizes wake interaction, all simultaneously. A typical modeling comparison looks like this (indicative, project-dependent):

Export: HVAC vs HVDC, and where the crossover sits

Beyond the array, the project must move full array output to an onshore grid coupling point. Greater Changhua’s phases connect via onshore substations in Taiwan’s Changhua County with high-voltage alternating-current (HVAC) export that lands around 161 kV / 345 kV class coupling — representative of the central-western Taiwan grid. The engineering choice is between HVAC export and high-voltage direct current (HVDC):

• HVAC is simpler, cheaper at moderate distance, and handles reactive power with shunt compensation, but its cable charging current consumes ampacity in proportion to distance. Beyond roughly 60–80 km of submarine cable, the charging-current penalty makes HVAC increasingly lossy and expensive to compensate.
• HVDC decouples active-power transfer from cable charging, eliminating the distance limit, but at the cost of two converter stations and their losses (typically ~1–2% total, split across rectifier and inverter) plus far higher substation cost and complexity.

For a Taiwan Strait project 35–50 km offshore, HVAC with reactive compensation is comfortably inside the economic envelope — why the transmission choice tends to be converter-light at this range. But the modeling discipline is identical for both: model the export path’s losses, reactive compensation, and voltage profile jointly with the array, not after it. The crossover distance where HVDC wins shifts with array rating, cable type, and the cost of losses at the grid coupling point — one more optimization that a fixed-percentage model hides.

Grid Connection and Wind-to-Grid Technical Compliance

Commissioning 1 GW of offshore wind onto Taiwan’s grid is as much a grid-code exercise as a wind-farm exercise. Taiwan’s grid code for wind — enforced at the transmission coupling point — requires the farm to behave like a conventional power plant across four dimensions a pure yield model ignores:

• Fault ride-through: the array must stay connected through low-voltage (and in many modern codes, high-voltage) grid disturbances, typically down to zero voltage for a defined clearing window, and support recovery afterward. Because offshore plants are often electrically “remote,” this is validated with detailed converter-level phasor and EMT studies — not static percentage losses.
• Active-power and ramp control: the plant must be able to cap output and limit ramp rates on operator command. In a 1 GW plant that is a real energy-management function — and it is exactly where curtailment in the AEP budget originates. Modeling must represent how often and how deep curtailment actually bites, which depends on grid capacity and market signals, not on a flat annual percentage.
• Reactive-power / voltage control: offshore plants are increasingly required to supply regulated reactive power and hold voltage at the point of connection. Wind farm “grid-forming”-type and advanced P/Q capability changes the compensation design and interacts with the export cable’s charging requirements.
• Frequency response: converters must support inertia-like and fast frequency response in systems with rising inverter penetration. Taiwan’s high wind shares during the winter monsoon make the frequency response characteristic of a 1 GW offshore plant genuinely dispatch-relevant on many days.

For the modeler the practical translation is that “wind-to-grid” compliance produces operational constraints that feed back into AEP and revenue. A turbine fleet rated for a 48% gross capacity factor may be capped, ramped, or curbed during parts of the monsoon season precisely because the wind is too good for the grid to absorb. The reputable way to size the loss is to run plant-level dispatch against grid and market constraints — not to assume a single curtailment number. This is the same coupling between physics and operation that BESS projects face, and the reason I point wind developers at best-practice simulation methodology rather than static-loss shortcuts.

Modeling a Changhua-Scale Array Properly

Putting all of this together, a defensible GW-scale offshore model has to be hourly, loss-interacting, and grid-aware — three properties a spreadsheet multiplication chain generally lacks:

1. Hourly resource + power-curve stack. Apply the wind rose and turbulence to turbine power and thrust curves across all 8,760 annual hours, not a single capacity-factor times nameplate. Energy Optima’s wind module pairs manufacturer turbine power curves from a database of 118+ wind turbines against hourly resource series, so wake and availability losses land on a real time-series, not an annual scalar.
2. Losses that interact with the layout and the grid. Electrical loss should fall out of the actual array-cable geometry and voltage choice you model; wake loss should be a layout parameter you sweep, as in figure 1; curtailment should come from a dispatch run against a grid/generation mix where you can see when the monsoon wind overwhelms absorption. That is exactly the interaction structure an LP-based dispatch framework exists to capture.
3. Financialization of the loss stack. A GWh lost to wake is worth something different from a GWh lost to curtailment — the former was always going to be sold, the latter may be shifted by coupling with storage or firmed differently. Running the net AEP through a 25-year financial projection with realistic degradation of the underlying plant makes the array-level engineering choices trade-able against their cost, which is how you actually optimize a 1 GW array rather than “design” it once.

The Greater Changhua 2b & 4 commissioning is a useful forcing function for the whole industry’s modeling maturity, because it is the kind of project where the “easy” 1–2% errors — a mis-tuned wake model, a fixed curtailment scalar, an electrical loss pulled from a table instead of the cable geometry — each cost real annual GWh on a genuinely world-class resource. In the Taiwan Strait the wind is not the constraint on what a farm delivers. The array design, the export path, and the grid are. Those are all modelable — but only with a tool that treats them as interacting variables rather than static percentages.

For a working contrast in the offshore-vs-grid trade space at single-project scale, our He Dreiht analysis looked at 15 MW turbine economics in the North Sea; the same turbine-database-plus-hourly-resource approach scales cleanly to a Strait array built from 9–11 MW class machines.