We priced curtailment wrong for six years. Here is the model we use now.

Every constraint forecast we bought treated curtailment as a haircut on annual yield. It is not — it arrives in a handful of hours, correlated with exactly the hours the asset was supposed to earn.

We paid for six of those forecasts between 2019 and 2024. Each one produced a single percentage: 3%, 5%, 8% of annual generation lost to constraint. We put that number into the model as a flat reduction on output and moved on. Two of the assets underwrote on that basis have never made their case, and the reason is not that the percentage was wrong. On one of them it was almost exactly right.

What we measured

In 2024 we bought three years of half-hourly settlement data for eleven of our own sites and re-derived the number ourselves. The annual percentages came back close to the forecasts — within 1.4 points on nine of the eleven. The distribution did not.

• At the worst node, 71% of the curtailed energy fell in 4% of the hours.
• Those hours were not random. They clustered in the same weeks, and within those weeks in the same part of the day.
• They correlated at 0.62 with the hours in which the day-ahead price was in its top decile.

That last line is the whole article. Curtailment does not remove average megawatt-hours. It removes the expensive ones — the hours a wind asset earns most of its margin in, because the same weather that fills the network is the weather that fills your rotor.

Why a flat percentage flatters the case

If you take 8% off annual output and price the remainder at the annual capture price, you have priced the loss at the average. The energy you actually lost was worth substantially more than the average. On the eleven sites we re-derived, the revenue effect of curtailment was between 1.3 and 2.1 times the volume effect. The median was 1.6.

Applied to a 90 MW onshore project we underwrote in 2021, that is the difference between a forecast that clears its hurdle rate and one that does not. We built it. It has cleared its cost of debt every year and has never once hit the equity case.

A forecast that gets the annual number right and the shape wrong is not 90% correct. It is a different forecast about a different asset.

The model we use now

We stopped buying single percentages. What we ask for, and what we now build ourselves where the data exists, is an hourly constraint profile aligned to an hourly price curve — and we run the asset against both together rather than netting one off the other.

In practice that means four things.

• Half-hourly, not annual. Three years minimum at the specific substation, not the zone. Two nodes on the same zone in our own portfolio differ by 6.8 points.
• Curtailment and price from the same hours. If the two datasets are not aligned to the same timestamps, the correlation that matters is invisible by construction.
• A separate downside case for the shape. We hold the annual percentage flat and re-shape the losses into the top price decile. That is not a stress test of the forecast; it is a stress test of our own assumption that the forecast is describing what we think it describes.
• Explicit treatment of compensation regimes. Where constraint is paid, the shape matters less. Where it is not, it matters more than anything else in the model. Two of our markets changed regime inside the holding period.

What it changed

Re-running the eleven sites through this model moved four of them across the line we use to decide whether to proceed — three downwards, one upwards. The one that moved up was a storage-adjacent project where the correlation ran the other way and the constraint hours were hours we could buy in.

It also changed what we do at the development stage. A queue position at a node with 11% curtailment concentrated in winter evenings is a different asset from one with 11% spread evenly, and we now price the two differently at the point where we decide whether to spend money on land.

What we would still like to be better at

We do not have a defensible way to forecast regime change. Two of the eleven sites sit in markets where the compensation rules moved after financial close, in one case favourably and in one case not. We model both directions and we do not pretend the resulting range is a probability.

We are also aware that three years of history at a node is a small sample for an effect that is driven by network build-out on a ten-year cycle. Where a reinforcement is scheduled inside the holding period, the historical shape is evidence about the wrong network.

If you are buying a constraint forecast this year, the question worth asking the provider is not what the percentage is. It is: what does the loss look like hour by hour, and can you show it against price for the same hours? On four of the six we bought, the answer was that the underlying model did not produce that output at all.