Effectiveness of an artificial intelligence-based system to curtail wind turbines to reduce eagle collisions

Abstract

Operating wind-power projects often includes protecting volant wildlife. One method for doing this uses an automated system to detect, identify (through use of artificial intelligence; AI), track animals (targets) and curtail turbines when risk of a collision is high. However, assessments of the effectiveness, in terms of identification accuracy and subsequent turbine curtailment of such systems are lacking. Over 1 year, we assessed such an automated system installed at a wind project in California, USA to determine its identification accuracy and rates at which "virtual” curtailments were ordered (without slowing turbines), for eagles (intended targets) and non-eagle targets. The system correctly identified 77% of eagles and 85% of non-eagles. Curtailment orders occurred 6 times more frequently for non-eagle targets (5,439) than for eagle targets (850). Greater abundance of common ravens that were misidentified as eagles influenced the effectiveness of the system by greatly increasing unintended curtailment orders. The balance between costs (price of the IdentiFlight system, reduced energy generation, turbine wear and maintenance) and benefits (reduced collisions between intended target species and turbines) may depend upon the biological setting, speed at which operators can curtail turbines, and the objectives of the operator when considering the IdentiFlight system.

Variation in flight characteristics associated with entry by eagles into rotor‐swept zones of wind turbines

“…For example, we examined the flight paths of eagles that triggered curtailment, but the act of curtailment itself may affect an eagle's flight behaviour. Further, the machine-vision system used at our study site had a false classification rate for non-eagles of 28% in a proof-of-performance test (McClure et al 2018), and the classification algorithm appears not to have improved over time (Duerr et al 2023). Some unknown proportion of the flight paths that we analysed were therefore of non-eagles.…”

“…2018), and the classification algorithm appears not to have improved over time (Duerr et al . 2023). Some unknown proportion of the flight paths that we analysed were therefore of non‐eagles.…”

Methodological Selection of Optimal Features for Object Classification Based on Stereovision System

“…Much effort has been recently made to develop solutions that ensure comprehensive protection of birds and meet the challenging requirements of environmental authorities [ 8 , 9 ]. The bird protection system (BPS) uses stereovision that provides precise 3D localisation to prevent collisions with wind turbine blades [ 10 , 11 ]. However, reliable species classification can improve deterrence systems and reduce wind turbine stoppages [ 12 , 13 ].…”

“…However, reliable species classification can improve deterrence systems and reduce wind turbine stoppages [ 12 , 13 ]. Hence, a reliable real-time bird species/genus/family classification algorithm recently became an emerging research topic [ 11 ].…”