From bird collision risk to curtailment decisions: where UK wind projects are heading

At Balmeanach Wind Farm on Skye, the collision risk model still predicted 37 white-tailed eagle deaths over the project’s 40-year life, even after one turbine was removed to reduce the risk. Removing it had lowered the modelled risk by 0.873 birds a year, leaving a residual prediction of 0.944 fatalities a year. In response, NatureScot recommended a package of monitoring and mitigation that included automated or observer-led shutdown on demand, as well as scheduled curtailment during periods of elevated collision risk.

That recommendation is where the series of our articles has been heading. In Part 1, we looked at how small a sample of bird activity vantage point surveys actually capture. In Part 2, we followed that sample into the collision risk model and saw how a modelled risk can be challenged by reality: three eagles killed by turbines in Donegal at sites consented with lower predicted impacts, and at Aberdeen Bay, a forecast of up to 8.54 collisions per turbine a year against none observed in 19 months. Shutdown on demand is a sensible response to that uncertainty, but it turns curtailment into a question that affects a wind farm throughout its operational life: how well can a bird’s presence be distinguished from an actual collision risk? Getting that distinction wrong in either direction has consequences: risk can be underestimated, leaving vulnerable birds exposed, or overestimated, leading to unnecessary production losses.

How shutdown on demand decides when to stop

The Band model is a modelling tool, not a measurement tool

The Band model estimates how many birds are likely to collide with turbines by combining sampled flight activity with assumptions about flight height, avoidance behaviour and turbine operation. This distinction becomes particularly important when a modelled risk begins to determine how turbines operate. NatureScot does not present the resulting collision number as an exact observation. Its methodology includes a separate Stage F specifically for expressing uncertainty from sampled flight activity, missing observations during conditions such as night and storms, imprecise flight-height estimates, variability in bird behaviour and physical characteristics, and simplifications within the collision model itself. The result is intended to be presented as a best estimate surrounded by an uncertainty range.

Bird detection doesn’t always equal collision risk

As that uncertainty carries into operations, the industry has several ways of managing it. Turbines can be curtailed during predefined seasons, times of day or conditions associated with elevated risk. Automated shutdown on demand makes that response more targeted by reacting to birds detected in real time. But the next step is not simply more automation, it is improving the intelligence layer between detection and intervention: distinguishing presence from actual risk, and understanding when, where and under what circumstances a particular flight is most likely to result in a collision.

Large empirical datasets are beginning to show what that distinction looks like in practice. A 2025 Journal of Applied Ecology study analysed 205,867 bird trajectories across 11 wind facilities and found that bird presence did not necessarily correspond with collision risk. The researchers distinguished exposure from sensitivity based on turbine proximity and time spent in the risk zone, showing that weather, season, time of day and flight behaviour all influenced risk.

Spoor’s monitoring of a floating offshore wind turbine at the SEM-REV test site in France shows how far presence and risk can sit apart. Over around ten months, a camera looking up at the blades produced 1,690 bird tracks confirmed by expert review. Four in five of those birds passed through without interacting with the turbine at all, and most of the rest avoided it by adjusting course at a distance rather than close to the blades. Two collisions were recorded, both gulls, equivalent to 0.1% of the tracks analysed. The site is offshore and the birds mostly seabirds, but the pattern matters for any curtailment decision: every one of those birds was present at the turbine, yet only a small fraction of flights came close to the blades, and a stop triggered by presence alone would treat them all the same.

For curtailment, then, the deciding evidence should not simply be whether a bird is present, but what we know about the risk represented by that particular flight at that particular moment.

How shutdown on demand decides when to stop

Intelligence layer behind the ‘stop’ command

An automated system does not simply detect a bird and stop a turbine. Between those events sits a decision chain:

A bird one kilometre away and flying off is not the same as one approaching rotor height, and a bird passing through is not the same as one foraging around a turbine. A system can be extremely good at detecting birds and still leave an important question unanswered: which detected flights actually have a high risk of collision? That question is what separates shutdown on demand from smart curtailment. Shutdown on demand reacts to a bird being detected; smart curtailment assesses the risk of that particular flight, using its trajectory, speed and flight height together with weather and turbine conditions, and stops the turbine only when that risk justifies it.

Balancing bird protection and the cost of curtailment

That balance matters to both sides of the equation. Statutory nature bodies and NGOs need confidence that vulnerable birds are protected. Developers need confidence that operational constraints are proportionate to the risks they are intended to mitigate. A 2026 review of 52 collision-mitigation approaches found a general lack of data on the energy-production losses associated with mitigation and specifically called for better evidence to allow conservation benefits to be balanced against the economic cost of curtailment.

As shutdown on demand becomes a more visible part of the mitigation conversation in the UK, detecting a bird early enough to stop a turbine is only half the task. The other half is the question this article began with: how good are we at deciding when that turbine actually needs to stop? The intelligence layer that makes that decision deserves the same evidence and validation as the collision model that led us there.

How good is the evidence behind your curtailment or shutdown-on-demand decisions? If you’d like to look at that for your site, contact us.