Pressures to decarbonize the United States’ electricity production, reduce dependence on foreign energy imports, and the declining levelized cost of renewable electricity is making wind energy an increasingly appealing means of meeting electricity demand in the United States. However, the installation of new commercial wind farms to meet this demand requires knowledge of the most suitable locations for their installation, which depends on a combination of environmental, technical, economic, political, and social characteristics. Wind Farm Site Suitability (WiFSS) models are frequently enlisted to assist in this decision-making process in countries around the world for both onshore and offshore wind farm siting decisions. However, existing WiFSS models serve to assess present-day wind farm siting potential, rather than project specific locations for future wind energy development. Taking cues from Socio-Environmental Systems (SES) models of urban growth, this dissertation presents a Logistic Regression-Cellular Automata (LRCA) model, henceforth referred to as WiFSS-LRCA, conceived to produce maps that identify scenarios of potential future locations and timing of future commercial wind farms across the Conterminous United States (CONUS) between now and the year 2050.

Following a review of existing WiFSS modeling approaches, and of common practices by which WiFSS modeling studies select and represent their predictors, the niche that WiFSS-LRCA serves to fill was consequently identified. The majority of WiFSS studies take a Geographic Information Systems-based Multi-Criteria Decision Analysis (GIS-MCDA) approach that combines spatial data layers corresponding to selected predictors to construct a composite suitability surface. Other common approaches include Non-GIS-MCDA models that rank discrete potential wind farm sites to prioritize their order of development, Bayesian Network (BN) models that construct and convey probabilistic relationships between predictors, and Logistic Regression (LR) models that perform either spatial or non-spatial assessment of a wind farm’s suitability of presence based on the log-odds of a linear combination of predictors. The common limitation of these modeling approaches is their lack of a temporal component, meaning that they can assess WiFSS only at a single point in time. WiFSS-LRCA fills this niche by combining an LR equation with the decision rules of Cellular Automata (CA) to iteratively advance the computed probabilities of each grid cell, based on areas constrained from development and neighboring grid cells that already contain wind farms.

WiFSS-LRCA enlists a large set of predictors ranging from wind speed to legislation in effect in order for the model to represent the influence that environmental, technical, economic, political, and social predictors have on wind farm siting decisions. Data were aggregated at 20 different grid cell resolutions, collated in four different predictor configurations, and adjustments to the model’s constraint, neighborhood effect, and equation-based scenario transition rules were incorporated into the model’s construction, facilitating WiFSS-LRCA’s sensitivity and scenario analysis of model outputs by end-users. WiFSS-LRCA incorporates both calibration of its LR equation’s predictors and validation of the model’s performance to determine its ability to correctly identify the observed locations of present-day wind farms. Subsequently, the model constructs a WiFSS map whose interpretation and predictive accuracy are informed by the calibration and validation process. Construction of scenarios that modify WiFSS-LRCA’s predictors allow for the model to consider the impacts of changes in these predictors on the locations of future wind energy development (e.g., new transmission line construction, opinions of wind energy improving with time, increasing temperatures due to climate change).

The ability of WiFSS-LRCA to produce suitability surfaces with verifiable accuracy is greatest under the following conditions: when running the model over an individual U.S. state rather than the CONUS, when using a smaller grid cell size, when using a more complete (Full configuration) or more refined (Reduced configuration) set of predictors, and when the selected study area contains a larger number of present-day commercial wind farms. Across most study areas, however, WiFSS-LRCA is typically able to correctly identify 75-85% of grid cells that do and do not contain commercial wind farms, with these classifications most often associated with high wind speed, proximity to transmission lines, legislation that supports wind energy development, and large tracts of undeveloped land. CONUS-level model runs indicate five regions as being the most suitable for present wind energy development: Southern California, the Pacific Northwest, the Central Plains, the Great Lakes, and the Northeastern United States. CONUS-level model runs have a tendency to over(under)-estimate grid cell probabilities within (outside) the Central Plains and Great Lakes, which makes state-level model runs useful for revealing smaller-scale differences in the probabilities computed within these five broad regions.

Subsequent iterations of WiFSS-LRCA out to the year 2050 show projected wind energy development to remain concentrated within these same regions. Many of the grid cells initially classified as false positive in the model’s first iteration are those that gain wind farms in subsequent iterations, particularly false positive grid cells that were part of high-probability hotspots identified by Getis-Ord statistics. Running WiFSS-LRCA over states outside of these five regions projects wind energy development potential in low-probability areas (as shown in this dissertation for Florida and Kentucky) with projected wind farms in these states concentrated closer to existing infrastructure and away from protected natural areas. The Odds Ratios (ORs) computed during WiFSS-LRCA’s initial calibration provide geographical insight into its projections, with grid cells characterized by high wind speed, undeveloped land, and ambitious Renewable Portfolio Standards (RPS) being the most likely to gain wind farms in future decades. The model’s projections are, however, shown to be sensitive to end-user definitions of parameters, with neighborhood effect and constraint definitions greatly affecting the location and timing of projected wind farm locations. The scenario setup, by contrast, is shown to mostly influence the timing of these projections, with grid cell size moderately affecting both.

Multiple limitations exist in the application and interpretation of WiFSS-LRCA. Firstly, the lack of existing LRCA approaches to assessing wind farm siting potential meant few standards existed to guide this model’s development, such as the setting of default constraints and establishing cutoff statistics for refining the model’s enlisted predictors. Secondly, the use of an LR equation to construct suitability surfaces in the model’s first iteration means that both classes of the dependent variable must be filled, requiring a study area to contain at least two commercial wind farms, compromising the model’s reliability in runs over the Southeastern United States. Finally, the lack of spatial stratification during WiFSS-LRCA’s calibration and validation means that the model is trained to recognize predictors associated with wind energy development in regions where many wind farms exist, namely the Central Plains and Great Lakes, hence the greater number of Type 2 errors in CONUS-level model runs outside of these regions. Selecting stratified samples of grid cells that contain wind farms from different parts of the CONUS could be incorporated into WiFSS-LRCA to address this bias. Other directions for future work with WiFSS-LRCA include the following: optimization to assess offshore wind energy development potential by training the model with proposed offshore wind farm sites surrounding the CONUS; adapting WiFSS-LRCA to run over multiple states simultaneously to identify predictors that influence wind farm siting decisions at regional spatial scales; and performing projections of other types decentralized land-use change, such as solar energy development given similarities in the required model predictors.