Abstract
The primary objective of this study was to explore the use of regional-scale, i.e., synoptic, meteorological variables to predict broader-scale movements of bats to further the goal of reducing bat fatalities at wind energy facilities. This effort represented a preliminary, proof-of-concept approach for future exploration, and its intent was not for operational use in the near-term. We investigated the feasibility of decision support tools for bat fatality reduction that employ machine learning approaches, more specifically gradient-boosted linear regression trees. We modeled bat mortality and bat activity, estimated as bat echolocations derived from audio recordings of echolocation activity, with 31 predictor variables including: atmospheric conditions at surface and pressure-level measures of wind speed; air temperature and turbulent kinetic energy (a measure of wind velocity fluctuation); landcover characteristics (USGS 2011 National Landcover Database, 30 m resolution) at 100 m and 2 km surrounding acoustic recording sites; temporal data (e.g. ordinal date); and geographic data (e.g. latitude). The intent of our modeling effort was to produce a model for forecasting when, where, and under what conditions bats would occur in close proximities to wind turbines.
We explored the potential application of the most current modeling methodologies for forecasting bird migration based on weather variables (e.g., BirdCast) for predicting bat acoustic activity and collision risk. We chose to focus this initial effort on the Great Lakes region because of the density of available data and accessibility with respect to bat distribution information, existing mortality and acoustic data, and predictor variables, with a plan to develop additional region-specific models for bats for future testing at multiple wind facilities. We used gradient-boosted regression trees to generate two models: 1) a regional model using acoustic recordings of echolocation activity in the Great Lakes region to predict bat echolocation based on weather, habitat, and geographic variables; and 2) a broader but more diffuse model with samples from bat fatalities at wind energy facilities to predict periods of heightened bat collision risk from weather and geographic variables. At present, there is no model to directly relate preor post-construction acoustic activity and bat fatalities at wind facilities. To evaluate the efficacy of using weather variables with the power embodied in the BirdCast migration forecasts that are extensible across the contiguous United States, we separately modeled these distinct and different data representing bats.
The best all-species gradient-boosted regression tree model explained 73.0% of the variation (R2=0.73) in bat echolocation activity from acoustic data from nearly 120,000 recording periods and produced explanatory predictions for areas that were not directly sampled. This acoustic activity model was based on the following most important variables whose contributions we determined in terms of their gain in predictive performance: 1) surface temperature; 2) % water within a 2 km radius; 3) hour after sunset; 4) ordinal date; and 5) hour before sunrise. The best all-species mortality model explained 44% of the variation (R2=0.44) in bat fatality data from post-construction fatality monitoring datasets, with the most important variables being 1) longitude, 2) ordinal date, and 3) latitude; followed by 4) surface, and 5) air temperatures. We did not include landcover predictors in the mortality model and did not have a sufficiently broad geographic and spatial sampling of fatality search data (6,289 site-search days; 226,842 turbine-search days), or the necessary randomization in methods, to apply this model to areas beyond locations with data – i.e., we could not expand to novel spatial and temporal extents because of a lack of fatality search data.
Future research should clarify how predicted acoustic activity relates to mortality and at what geographic scales. Model applications, including use of weather, habitat, and spatio-temporal variables to predict bat activity and fatalities in places where monitoring may be limited, are still possible without assurances of a strong relationship with fatality risk. Modeling for broader scale applications – e.g., greater geographic extents, larger temporal periods, more sites and locations – requires randomly sampled data over greater areas and longer periods of time (more hours of night, more nights per season, more years), while maintaining high spatial and temporal resolution. Results from our efforts highlight some guidance with respect to potential expectations of model behavior, parameters, and accuracy, as well as important caveats regarding availability and robustness for statistical efforts with existing data.
We produced proof-of-concept, generalizable models that predicted heightened bat echolocation activity and number of fatalities from variables, including habitat type, atmospheric conditions, and indicators of seasonality and geography. We used bat echolocation as a surrogate for bat activity, and bat collision risk was estimated from bat fatalities collected during carcass searches. These two independent models were spatially explicit in terms of their data inputs, predicting at a localized scale in their respective study areas. Acoustic data were generally more accessible than mortality data, primarily because of experience, expense, and availability of acoustic monitoring methods relative to searching for carcasses. The models performed well, and results support further exploration to expand analysis with data from regions beyond those encompassed by this study and could be considered for integration with smart curtailment algorithms. Such consideration of application in operational environments will require additional research guided by this proof-of-concept study. Our results complement efforts to increase the efficiency of bat risk minimization strategies focused on individual projects, by providing a tool for predicting movements or activity of bats and thus bat collision risk over larger spatial scales during migration, movements to or from maternity roosts and hibernacula, or mass roost emergences (as witnessed in Mexican free-tailed bats). Our efforts support further exploration of smart curtailment algorithms that account for weather variables at a regional scale as opposed to relying on weather data from met towers located within projects. Regional models could provide advanced warning of periods of increased risk to bats and allow energy producers and managers of regional electric grids to plan for periods of reduced power generation.