Morlais’s tidal project will use AI to separate marine wildlife from drifting debris

A tidal-energy installation needs to distinguish a seal from seaweed under conditions far less orderly than a laboratory image collection. HydroWing, Inyanga Marine Energy Group’s tidal division, has contracted Ecodetect to provide AI-driven wildlife monitoring for its planned 30 MW project at Morlais in Wales. The agreement was announced on 27 October.[1]

The proposed system combines underwater cameras and sonar in surface- and seabed-mounted equipment. Ecodetect says it will classify marine animals separately from debris, algae and seaweed, then relay the results so operators can trigger environmental mitigation. The announcement describes an operator response, not a demonstrated autonomous turbine-shutdown system.[1]

The first HydroWing device is scheduled for 2027, with 18 devices planned by completion in 2030. The announcement does not supply independently verified detection accuracy or response-time results. It is evidence of a procurement decision and a planned operational role, not yet evidence that the system has reduced wildlife harm.[1]

Thirty megawatts is not the whole site

Morlais is a larger shared-infrastructure scheme with consent for up to 240 MW. The 30 MW HydroWing project is one development within that setting. Keeping those figures separate matters: a monitoring contract for one developer should not be reported as deployment across the entire consented capacity.[1][2]

The project also has an environmental history that predates this AI contract. Morlais reported in September 2024 that Natural Resources Wales had discharged a condition relating to its environmental monitoring and mitigation plan. It described baseline information and the role of monitoring in development. The new classification system enters an existing framework; it does not replace the need to establish what is happening around the equipment.[3]

An alert has two jobs

The first job is perception: decide what an observation probably contains. The second is operational: deliver information sufficiently clear and timely for somebody to act. A classifier could perform well on stored clips yet be much less useful if its results arrive late, cannot distinguish uncertainty from absence, or are difficult to connect to a specific location.

False alarms and missed detections create different problems. An unnecessary intervention could interrupt generation; a missed animal could undermine the protection the system is intended to support. That trade-off cannot be assessed from a single attractive accuracy percentage. The relevant evidence would include the types of mistakes, the conditions in which they occur and the action associated with each alert.

For example, a hypothetical alert saying “possible marine mammal near device three, low confidence” conveys a different decision from “fish-like object somewhere in the camera view.” A useful deployment evaluation would ask whether an operator can make an appropriate decision from the available information. It would also examine how the system behaves when visibility or communications deteriorate.

The potential value extends beyond an immediate response. A consistently recorded sequence of observations, alerts and actions could help researchers and operators understand interactions over time. That requires records that are interpretable and auditable, not simply a running total of detections.

The next milestone is therefore not just the installation of an AI box beside a turbine. It is a functioning chain from observation to classification, communication and response. The contract gives that chain a commercial starting point; the planned deployment must establish how well its links hold together.

Report an error or suggest a correction