Abstract
Collision of birds and bats with wind turbines is a conservation concern for both land-based and offshore wind projects. The fatality rates of birds and bats at land-based turbines are well documented. The measurement strategies on land focus on finding carcasses following collision, estimating the number of carcasses missed through searcher efficiency, carcass persistence trials and carcass fall distributions, and modeling statistically robust fatality rates. Few technologies have been developed to monitor offshore bird and bat collisions, and many that have been developed focused on detecting collisions with large birds. The few studies that have attempted to document collisions at offshore turbines do not account for smaller bodied animals or for collisions that might be missed, which prevents the calculation of statistically robust fatality rates. The overall goal of this report, A Multi-Sensor Approach for Measuring Bird and Bat Collisions with Offshore Wind Turbines (Project), was to develop an effective multi-sensor system for quantifying bird and bat collision rates, specifically for offshore wind facilities. The Project goal and resulting automated collision detection system was achieved through two major technological advancements: 1) refining The Netherlands Organisation for Applied Scientific Research’s (TNO’s) existing WT-Bird® vibration sensing system, that had successfully detected large bird collisions during daytime, to allow for improved detection of smaller birds and bats during both daytime and nighttime hours and 2) improving image processing systems and developing and integrating machine learning algorithms to automatically detect and classify small and large bird and bat collisions with offshore turbines.
This research effort used a sequential, step-wise approach to update and validate the WT-Bird® system with advanced collision detection capabilities and sensors. During Task 1, TNO updated the WT-Bird® system, including improved fiber optic sensors for use in the turbine blades and a computing system that could better differentiate collisions from turbine operational noise—as heard through the blade— and environmental noise, including rain and hail. Additionally, cameras were updated with 20-megapixel color models suitable for capturing an adequate number of pixels near rotor hub height (80 meters [m]), such that a machine learning (computer-vision) approach could be trained to identify birds or bats involved from imagery. Following the improvements to the WT-Bird® sensor system at the TNO laboratories, the updated WT-Bird® system was shipped to the National Renewable Energy Laboratory (NREL) National Wind Technology Center’s (NWTC’s) Flatirons Campus in Arvada, Colorado.
For Task 2, the updated WT-Bird® system was installed in a General Electric 1.5-megawatt wind turbine by March 31, 2021. The final installation configuration of sensors in the blades included one blade with three sensors (six m from blade stem, 12 m, 18 m), and two blades with two sensors (six m, 12 m). The sensor configurations were included to evaluate the efficacy of detection recognizing increased detection may require additional sensors. The NREL used a pneumatically controlled launcher with three size classes of projectiles (balsa wood frame and food-grade gelatin, mixed with coyote urine to deter scavengers) to test the system’s ability to detect collisions with objects of similar weight and size to birds and bats, referred to here as collision challenge trials. Projectile sizes included small (8 grams [g]; a small species, such as a warbler or bat), medium (25 g; a large sparrow or flycatcher), “middle” (40 g; grosbeak or oriole) and large (250 g; a gull or small duck). Collision challenge trials (39 trials for small projectiles, 42 for medium, 37 for middle, and 28 for large) were conducted while the turbine was operational and generating electricity by NREL engineers, who shot projectiles at the blades with a pneumatic launcher. Minimum collision detection rates were 62% for small and middle sized projectiles during the trial, and greater for larger objects. Detection rates were greater for blades with three sensors, rather than two, for all projectile size classes. The detected rates of collision for the blade with three sensors ranged between 0.65 and 0.75 detections per known collision, depending on object size.
The Task 2 collision trials at the NREL’s NWTC provided substantial documentation that the WTBird® collision sensors detect collisions with objects as small as eight g and that detection rates exceeded the minimum rates established in the Statement of Project Objectives (SOPO) for advancing development (at least 20% detection rate for small and medium objects, at least 50% for large objects). Furthermore, the rates of false detection were low. TNO adjusted the system’s algorithms to differentiate between noise typical of operating turbines and collisions after the collision challenge test. The revised algorithms (from Algorithm 1 to Algorithm 2) increased collision detection rate while reducing the false positive rate toward zero. WT-Bird® collision detection rates were better than what might be expected during a conventional ground-based carcass search, particularly when including consideration of carcass persistence. Therefore, the decision was made to proceed to the subsequent turbine field trials with WT-Bird® (Task 4).
One of the original goals of the project was to evaluate if the monitoring system could detect collisions via the vibration sensors, and at the same time, be able to identify what collided with turbine blades using video imagery collected just prior to the collision to identify what collided with the turbine blade. Task 3 focused specifically on addressing the second advancement objective, development and testing of an image processing system and integrating machine learning algorithms to automatically detect and classify small and large bird and bat images that were collected just prior to the time collisions were detected by vibration sensors. Imagery used to train video classification algorithms to differentiate between small birds, large birds, and bats were collected at one offshore island in Maine, and two locations in Minnesota. We used auxiliary information from bat acoustic detectors and direct field observations to identify periods when bats and birds were likely to be present in the area and selected imagery collected during these periods for review and object annotation using the Computer Vision Annotation Tool Version 2.0.0. We reviewed annotations for accuracy and exported them in Common Objects in Context Version 1.0 format prior to model training. We excluded unidentified flying objects, including insects and airplanes, from model training due to poor data quality and limited sample size, respectively. WEST conducted model training using the Pytorch Version 1.9.1 library in Python Version 3.8.12. The images from the annotation dataset were split between training and validation, with 80% used for training and 20% used for validation.
WEST reviewed over 300,000 images, and annotated 1,637 birds and bats from 781 images containing flying objects. The image recognition model performed well on the validation dataset and Average Precision (AP) values for all categories, meeting or exceeding the accuracy criteria specified in the SOPO. AP ranged from 0.71 for bats to 0.91 for large birds, and the overall model had a mean AP of 0.83. Confusion among categories was low with a majority of misclassification instances occurring between bats and bird categories. When including mis-categorized predictions, only 4% of birds and 10% of bats were not detected. Classification rates exceeded the SOPO requirements of 50% accuracy and the effort advanced to Task 4 for the in-field validation trial.
The Task 4 Validation of the WT-Bird® on a land-based turbine was initiated by July 2022 and compared how the WT-Bird® system performed against a traditional ground-based postconstruction mortality survey. The goals of Task 4 included: 1) validate the comparability of a fatality estimate, developed using data from the WT-Bird® collision detection system, to a standard fatality estimate from a typical land-based carcass search study, generated using the GenEst fatality estimator (a generalized estimator of fatality; Dalthorp et al. 2018) 2) confirm whether the computer-vision system, developed by WEST, could serve as a second, independent estimate of fatality rates while reducing the amount of video data retained. The study compared how the WT-Bird® system performed against a traditional ground-based post-construction mortality survey.
Task 4 followed the approved peer-reviewed study plan (WEST 2022). We initiated field deployment of the WT-Bird® system at the University of Minnesota Eolos Wind Turbine (University of Minnesota turbine) in Rosemount, Minnesota, a Clipper Liberty 2.5-MW turbine with an 80-m hub height with 96-m rotor diameter. The WT-Bird® system was installed by early July 2022, and throughout the subsequent months, standard fatality searching occurred three times per week. The WT-Bird® system was operational from August 13 – November 3, 2022, with the period prior to September 6 focused on tuning the WT-Bird® system to the University of Minnesota turbine. The WT-Bird® system was operational 83.7% of the time. Storm-associated outages caused the system to not operate during 16.3% of the survey period. During the survey period, the blade vibration sensors documented 15 collisions, with 13 occurring during twilight or in full darkness. Concurrent land-based carcass searches by WEST detected 13 carcasses (not adjusted for searcher efficiency, carcass persistence, or carcasses that fell outside of search plots). Cameras were operational for a portion of the study period. Similarly, among the eight collisions that occurred when both color and thermal cameras were operational, seven collisions included objects observed by the thermal cameras, including six documented collisions. The color cameras documented fewer collisions (n=3) during the same time period.
Using detection probability estimates calculated during collision challenge tests of small and large projectiles at the NREL facility, while accounting for system down time, the corresponding lower and upper adjusted fatality estimates for birds and bats combined were 23.89 and 27.57 , respectively, for the study period. In comparison, the post-construction mortality field study yielded a GenEst fatality estimate of 16.8 birds and bats (90% confidence interval [CI]: 9.88–27.92 birds and bats). Thus, the estimates based on the WT-Bird® collision detections were within the 90% CI of the field estimate.
The monitoring system used to process and record video imagery during the field study at the University of Minnesota Eolos Wind Turbine failed due to overheating. The monitoring system was improved after the field study using a real-time edge image processor, and re-deployed during 2023 at the turbine for further testing. Redeployment of the real-time edge processing and monitoring system for 87 days in 2023 improved system operations as the system did not fail (overheat) or require intervention. A distinct computer vision model for the edge deployment allowed for faster data processing, saving 643 hours of video containing detected objects, which was a comparative 92.3% reduction in data storage.
The goal of Task 5 was to implement the WT‐Bird® on offshore turbines, but the team encountered repeated delays and challenges with obtaining final operator agreement to test WTBird® at a specific offshore facility within the grant time period. This challenge, in combination with depleted funds insufficient to support the remaining testing, halted subsequent Task 5 testing. This decision was made in coordination with the Office of Energy Efficiency and Renewable Energy team during February 2024.
This research and development effort documented successful improvement of the WT-Bird® collision detection system to detect small birds and bats, and WT-Bird® is the first collision detection system to validate results compared to land-based post-construction monitoring. The collision trials provide estimates of missed targets that can be used to estimate fatality rates, a significant improvement relative to other offshore collision monitoring systems. Advances were made in developing an edge-processing solution to reduce data storage requirements, which is important if the system is deployed for long periods of time at offshore turbines. The improved WT-Bird® system also provides an important option for wind operators on land or offshore who need to document specific details about when collisions occur, particularly efforts to further research on bat impact minimization, or when standard fatality searches are impractical (e.g. offshore) or inadequate (e.g. challenging locations on land).