You are on page 1of 10

APPLICATION NOTE

A Data Mining Approach to Estimating Rooftop Photovoltaic


Potential in the U.S.

Caleb Phillipsa , Ryan Elmoreb , Jenny Meliusc , Pieter Gagnonc , and Robert Margolisc
a
Computational Science Center, National Renewable Energy Laboratory, 15013 Denver West
Parkway, Golden, CO, USA; b Department of Business Information and Analytics, Daniels
College of Business, University of Denver, 2199 S. University Blvd., Denver, CO,
USAc Strategic Energy Analysis Center, National Renewable Energy Laboratory, 15013
Denver West Parkway, Golden, CO, USA

ARTICLE HISTORY
Compiled March 29, 2018

ABSTRACT
This paper aims to quantify the amount of suitable rooftop area for photovoltaic
(PV) energy generation in the continental United States (US). The approach is data-
driven, combining Geographic Information Systems (GIS) analysis of an extensive
dataset of Light Detection and Ranging (LiDAR) measurements collected by the
Department of Homeland Security (DHS) with a statistical model trained on these
same data. The model developed herein can predict the quantity of suitable roof area
where LiDAR data is not available. This analysis focuses on small buildings (1,000
to 5,000 square feet) which account for more than half of the total available rooftop
space in these data (58%) and demonstrate a greater variability in suitability com-
pared to larger buildings which are nearly all suitable for photovoltaic installations.
This paper presents new results characterizing the size, shape and suitability of US
rooftops with respect to photovoltaic installations. Overall 28% of small building
roofs appear suitable in the continental United States for rooftop solar. Nationally,
small building rooftops could accommodate an expected 731 GW of PV capacity
and generate 926 TWh/year of PV energy on 4,920 km2 of suitable rooftop space
which equates to 25% the current U.S. electricity sales.

KEYWORDS
Applied statistics; regression; GIS; solar; energy; predictive model

1. Introduction

Rooftops provide a large expanse of untapped area for solar energy generation using
photovoltaic (PV) technologies that convert solar energy to electrical energy. Onsite
distributed power generation could potentially reduce the costs and losses associated
with the transmission and distribution of electricity. However, there are open questions
as to how much energy could be generated by rooftops on existing buildings and
whether some areas are more suitable for these technologies. This paper provides a
data-driven analysis of U.S. (national, state, and zip-code level) rooftop photovoltaic
suitability and technical electricity-generation potential for small buildings.
We use light detection and ranging (LiDAR) data, GIS methods, and statistical

To whom correspondence should be addressed: Ryan Elmore at Ryan.Elmore@du.edu

Pursuant to the DOE Public Access Plan, this document represents the authors' peer-reviewed, accepted
manuscript. The published version of the article is available from the relevant publisher.
Tilt & Azimuth
Calculation

Rooftop Plane Suitability


LiDAR Source Data Extraction Determination Suitable Rooftop Planes

Shading
Simulation

Figure 1. A database of suitable planes underlies the model development. LiDAR data is used to extract
rooftop plane information. Planes are classified as suitable if their tilt, azimuth (orientation) and shading is
appropriate for PV applications. These data are used directly for development of tilt, azimuth and plane area
models. Aggregated suitability data are used to develop a suitability model.

modeling to characterize trends in the suitability of rooftops for PV in 128 cities


nationwide – representing approximately 23% of U.S. buildings. These data are used as
the basis for predictive modeling to estimate capacity for the entire continental United
States. In this paper, we focus our modeling efforts on small buildings of between 1,001
and 5,000 square feet because they constitute more than half of the total rooftop area
in the LiDAR data (58%) and show the greatest variability in terms of suitability. By
comparison, larger buildings are almost universally suitable for installation of PV [11]
and [13].
A number of tools and methods now exist to estimate the solar energy potential
of a single home or building, and these tools are being actively developed to enable
firms to market solar systems to individual building owners. While there have been
some detailed city- and state-level estimates [2, 10, 20], there have only been a limited
number of national level assessments of rooftop PV technical potential to date [3, 4, 16].
Further, most of the previous estimates have relied on rough engineering heuristics in
estimating the fraction of rooftop area suitable for hosting PV systems.
This report provides a detailed, data-driven modeling effort for U.S. (national, state,
and zip-code level) small building rooftop PV suitability and technical electricity-
generation potential. This paper focuses on the statistical model and data mining at
the core of our efforts to characterize this potential. Additional analyses of LiDAR
data for individual cities, as well as a higher level discussion of policy and technology
implications and photovoltaic models are available in [11]. The results presented here
could benefit a broad spectrum of stakeholders including local municipalities, solar
energy researchers, planners, utility companies, and policymakers.

2. Data

In this section we discuss data used for determining what rooftop area is suitable for
hosting PV systems in a given region. We begin with roof plane information extracted
from LiDAR and building footprint data to construct a “ground truth” dataset. Figure
1 provides a schematic of the data preparation. Data are pre-processed to determine
the shading, tilt, and azimuth of each rooftop at a horizontal resolution of 1 m2 . These
results can then be aggregated to determine the total quantity of rooftop area suitable
for PV systems at the building, zip code, utility service territory, state, or national
levels of resolution. In those areas where LiDAR data may not be available, we can
predict the amount of rooftop space with a model trained on this dataset. Hence, the

Pursuant to the DOE Public Access Plan, this document represents the authors' peer-reviewed, accepted
manuscript. The published version of the article is available from the relevant publisher.
Figure 2. Coverage of LiDAR data used in this study accounts for 128 cities and 40% of the U.S. population.

combination of predictive modeling and the LiDAR data enables nationwide estimates
at arbitrary scales.
The LiDAR data used in our analysis were obtained from the U.S. Department of
Homeland Security (DHS) Homeland Security Infrastructure Program (HSIP), and
was gathered between 2006 and 2014. These data are available exclusively to govern-
ment agencies upon request. For each of the 128 cities in the dataset, DHS provided
LiDAR data in raster format at 1 m2 resolution and a corresponding polygon shape-
file of building footprints. The raster data, which provide measurements on a uniform
grid, are based on the reflective surface return of LiDAR data wherein each pixel’s
value is the measured height of the structure (or ground) at that location. The DHS
dataset also includes detailed data for about 26.9 million buildings and 7.7 billion
square meters of rooftop area, or about 23% of U.S. buildings [5, 6]. The area covered,
shown in Figure 2, represents about 122 million people or 40% of the U.S. population.
Our first step in processing the LiDAR data is to run a shading simulation on the
digital surface model of each city. The standard ArcGIS Hillshade tool was used to
create a hillshade for every hour of daylight for March 21, June 21, September 21,
and December 21 [8]. The altitude (degrees above vertical) and azimuth (degrees from
north) of the sun is recorded for each individual location for every hour and input into
the tool. This results in approximately 45-50 separate hillshade files for every city. The
hillshade tool produces a range of values from 0 to 254 to approximate the amount of
illumination each cell receives. The hillshade files are reclassified into a binary dataset
to create a simple yes/no determination of whether the cell is in sun or in shade. Each
month was determined to have a different threshold of illumination required to assign
the cell as being in sunlight. March requires 60% illumination (values 153-254); June
requires 70% illumination (values 178.5-254); September requires 60% illumination
(values 153-254); and December requires 50% illumination (values 127.5-254). The
hourly reclassified hillshade rasters were summed to create four files, one for each day
modeled. They were also summed all together to create a simulated annual average

Pursuant to the DOE Public Access Plan, this document represents the authors' peer-reviewed, accepted
manuscript. The published version of the article is available from the relevant publisher.
hillshade value. We then use this metric to exclude roof area that is excessively shaded.
The orientation of a roof plane is important, both for determining whether it is
suitable for PV as well as for simulating the productivity of modules installed on the
area. As the first component of orientation, we determined the tilt of each square meter
of roof area within our LiDAR dataset. We define all roof area with a tilt less than 9.5
degrees as “flat”. This value was chosen to be consistent with many roofers’ and PV
installers’ definition of flat roofs for installation purposes. As the second component
of roof plane orientation, we determined the azimuth (aspect) of each square meter
of roof area. Each square meter is categorized into one of nine azimuth classes, where
tilted roof areas were assigned one of the eight cardinal directions and area with a
tilt less than 9.5 degrees was classified as flat. The azimuth values were then used to
identify roof planes by assuming contiguous areas of identical azimuth class were a
unique plane and aggregating each of the individual square meters of roof area into
polygons representing contiguous roof planes. For each of the individual roof planes,
the ArcGIS Zonal Mean tool was applied to the tilt raster to determine the roof plane?s
mean tilt [7]. The dataset produced through this process consisted of a raster giving
a single tilt value for each unique roof plane. All roof planes facing northwest through
northeast (292.5 to 67.5 degrees) were considered unsuitable for PV and excluded.
All tilt values greater than 60 degrees were removed from the dataset based on the
recommendation of PV installers.
The ArcGIS combine tool was used to create one large raster showing the final
hillshade, slope, and aspect values for all cells in the dataset. The combined file was
then filtered to identify only rooftop areas that met specific criteria for sunlight, slope,
and aspect. The suitable hillshade value was determined on a regional basis using
NREL’s System Advisor Model (SAM) tool [18]. For each city individually, SAM
calculated the number of hours a rooftop would need to be in sunlight in order to
produce 80% energy generation, taking into account weather data and the strength
of the solar resource. All roof planes facing east through west (67.5-292.5 degrees;
aspect values 3-7), all slopes less than 60 degrees, and all hillshade values above the
SAM-determined threshold were considered suitable for PV.
We defined building-level PV suitability as having at least one contiguous plane with
a projected horizontal footprint greater than 10 m2 that also met the shading, tilt, and
azimuth criteria. This sets the minimum area at ten contiguous square meters, which
provides sufficient area to install a 1.5 kW system assuming a 15%-efficient panel.
This minimum size threshold was selected to represent a conservative lower bound of
viable PV system sizes, based on current PV performance and historical patterns in
reported PV system sizing. Specifically, we reviewed reported system sizes for small
PV systems (< 10kW ) through 2013 [1] and determined that 96% of systems in this
class were greater than 1.5 kW.
The final dataset contains detailed footprint and suitable roof plane information
with both orientation and azimuth for all buildings in the area with LiDAR data. Each
suitable plane in the combined dataset is stored in a PostGIS database for extraction
and analysis [15]. These data are used to build a model, described in the next section
which aims to capture the key dynamics of the suitable rooftop characteristics.

3. Methods

In this section we describe a system of independent models that can be combined in


order to describe the key dynamics that govern the size and shape of rooftop area

Pursuant to the DOE Public Access Plan, this document represents the authors' peer-reviewed, accepted
manuscript. The published version of the article is available from the relevant publisher.
suitable for PV development. This system is useful for understanding the underlying
processes themselves and how they contribute to rooftop suitability for PV, as well
as for predicting capacity in those regions where LiDAR data is not available. The
three major components are rooftop suitability, rooftop plane area, and azimuth-tilt
distribution. When combined in an appropriate way, these models are used to predict
the average (and upper and lower estimates) for the amount of suitable rooftop area
available in the United States. Hence, we refer to the final model as our “composite”
model.
Previous estimates of U.S. rooftop PV technical potential focused on estimating
results at a regional or national level [4]. However, we find a significant amount of
variation in the percentage of small buildings that are suitable across the various
census regions and locale types. Thus, our predictive model leverages this variability,
along with additional attributes such as ground cover classification, population density,
and spatial characteristics, in order to allow predictions at a more precise resolution.
First, we describe the Small Building Suitability (SBS) model that we use to es-
timate pi , the percentage of buildings in area i that are suitable for rooftop rolar.
The independent variables considered in this SBS model are a combination of readily
available public demographic, census, and land cover data from sources including the
National Land Cover Database (NLCD 2006), U.S. National Center for Education
Statistics (2006), and U.S. Census Bureau. In particular, we considered the following
variables.

• Locale Description: National Center for Education Statistics (2006) locale


description, e.g., Small Town, Remote Rural, etc. of each zip code area.
• Census Region: The U.S. Census divides the country into nine regions: East
North Central, East South Central, West North Central, West South Central,
Middle Atlantic, South Atlantic, New England, Pacific, and Mountain.
• Land Cover Classification Percentage: NLCD (2006) percentages for each
zip code including Developed Open Space, Developed Low Intensity, Developed
Medium Intensity, Developed High Intensity, Forest, and Scrub.
• Population Density: 2010 population per square mile from the U.S. Census
Bureau was used.
• Location: northing (latitude) and easting (longitude)
• LiDAR Coverage: The percentage of a zip code covered by liDAR data.

While nationally the distribution of suitable rooftops appears skewed to the right,
zip code resolution small building suitability when grouped by region appears to be
normally distributed. This observation motivates our approach to fitting here, where
the geographic location (e.g., region and locale type) are included in the fitting process.
Our model building process consisted of constructing a multitude of models using
combinations of the variables mentioned above as candidate models for estimating
pi . We “trained” each model using a randomly selected subset of 75% of the original
LiDAR dataset and assessed its predictive accuracy using the remaining 25% of these
data by predicting the percentage of buildings that are suitable for PV and comparing
those predictions to their actual values. This standard cross-validation exercise was
repeated 10 times to produce averaged estimates of model performance. Ultimately,
our final model was chosen due to the fact that it was parsimonious and had a low
Akaike’s Information Criterion (AIC) relative to the other models. We note that the
proportion of predictions p̂i within 10% and 20% of the actual values pi using the final
model are 59.4% and 83.1%, respectively.

Pursuant to the DOE Public Access Plan, this document represents the authors' peer-reviewed, accepted
manuscript. The published version of the article is available from the relevant publisher.
Size of Plane Distribution Number of Planes Per Building
0.06
5e+05

Kernel−smoothed Density Function

Number of Occurances in Sample


4e+05

0.04

3e+05

2e+05
0.02

1e+05

0.00 0e+00

0 25 50 75 100 0 5 10 15
Per−Plane Area (m^2) Number of Planes

(a) (b)

Figure 3. Distributions of plane sizes and plane numbers for one million small buildings selected via simple
random sample from the LiDAR data. Figure (a) gives the kernel-smoothed probability density function that
describes the distribution of per plan area. Figure (b) provides the relative frequencies for the number of planes
on each building. Both appear to be exponentially distributed quantities.

The resulting SBS model1 is defined in Equation (1). For more details related to
our model building process, see [11].

pi = β0 + β1 Localei + β2 CensusRegioni + β12 Localei · CensusRegioni


(1)
+ β3 NLCD21i + β4 NLCD24i + β5 NLCD11i + β6 NLCD4·i + β7 Northingi + εi

where εi is assumed to be normally distributed with mean 0 and variance σ 2 . The


NLCD variables refer to land cover classes and, in particular, NLCD21 is developed
open space, NLCD24 is developed high intensity, NLDC11 is open water, and NLCD4·
is all types of forested land.
It should be noted that this model can produce values greater than 1 or less than 0.
In practice, we set any negative and positive predictions to 0 and 1, respectively. While
this solution is not as elegant as using, e.g., a beta regression model, our validation
has shown that it actually obtains better performance than a beta regression for the
same data. In practice this occurs very infrequently: 60 zip codes produce predictions
greater than 1 among the 31,904 total predictions. Future work may also consider
other regression methods designed for proportions, such a a generalized linear model
with logit link or an arcsin transformation of the response.
Next, we describe how to estimate quantity of suitable areas (planes) on rooftops,
and the total area available. In order to estimate these quantities, we took a simple
random sample of size N = 1, 000, 000 of small buildings across all zip codes where
LiDAR data is available, resulting in 1.86 million suitable planes. The resulting dis-
tributions for the size of plane and the number of planes from this sample are given
in Figures 3 (a) and (b), respectively.
We note that 99.6% of small buildings have six or fewer planes, and roughly 50%
have only one plane. We ignore rooftops having more than six planes. Buildings having

1 Fitted coefficients are too many to include in this paper, however predicted values for all U.S. zip codes can

be downloaded at [17].

Pursuant to the DOE Public Access Plan, this document represents the authors' peer-reviewed, accepted
manuscript. The published version of the article is available from the relevant publisher.
Distribution of Suitable Planes (Small Buildings)
Azimuth
E S SE SW W
2.0

1.5

15
1.0

0.5

0.0

2.0
Kernel Smoothed Density

1.5

28
1.0

Tilt (degrees)
0.5

0.0

2.0

1.5

41
1.0

0.5

0.0

2.0

1.5

54
1.0

0.5

0.0
0.001 0.100 0.001 0.100 0.001 0.100 0.001 0.100 0.001 0.100
Area (m^2) (Log Scale)

Figure 4. Estimated probability distribution functions (PDF) for each combination of azimuth and tilt for
small buildings.

one plane generally have a larger single plane and the per-plane area is similar for
buildings with two to six planes. Therefore, we categorize our planes into two classes
(one plane or more than one plane) and fit separate models to the classes in order
to estimate the average plane size. As can be seen in Figure 3, the size of plane
distribution appears exponential. We fit exponential distributions to the two separate
classes and found the average plane size for buildings with one plane and more than
one plane to be 33.8 m2 and 24.1 m2 , respectively. The overall average is 28.9 m2 .
Additional details and identical methods for medium and large buildings can be found
in [11].
The final piece to this composite model is in estimating the distribution of tilt
and azimuth orientations for the suitable planes on a given rooftop. We begin by
discretizing the azimuth and tilt information for each plane. The five azimuth bins
are 45 degrees each, covering 67.5 to 292.5 degrees. The four tilt bins equally split the
50.5 degree range from 9.5 degrees to 60 degrees. Anything less than 9.5 degrees is
considered flat. Most developable rooftop planes fall into the 28-degree tilt category,
and steep rooftops (54 degrees) are an order of magnitude less common than the next
category (41 degrees). Azimuths East, South, and West are most common, particularly
among the 28 degree category of rooftops. These azimuths correspond to buildings
aligned on a cardinal street grid.
Figure 4 shows the estimated probability density function for the fraction of planes
in each combination of tilt and azimuth for every zip code. This plot is log-scaled
and makes clear the long tail on each distribution and its respective lognormal shape.
In order to make estimates about the distribution of planes in unseen zipcodes, we
aim to estimate a matrix from these distributions T such that each element trc is the
average fraction of planes with that tilt (r = 1, 2, 3, 4) and azimuth (c = 1, 2, 3, 4, 5)
2
σrc
ˆ
combination, t̂rc = eµrc + 2 , or normalized by trc = P tP
rc
.
r c t̂rc
The complete, or composite, model is used to estimate the suitable area of rooftop
PV for an arbitrary geography. That is, let i be an geographical unit of interest. The
estimated total suitable area Ri for unit i is defined as follows. We first estimate the

Pursuant to the DOE Public Access Plan, this document represents the authors' peer-reviewed, accepted
manuscript. The published version of the article is available from the relevant publisher.
Figure 5. Average rooftop PV production per small building at the zip code level.

proportion of buildings that are suitable for PV by ni p̂i . We multiply this number
by the average rooftop plane area yielding ai . Finally, we distribute ai across the
tilt/azimuth matrix uniformly to obtain Ai = ai T (i) for the appropriate T defined for
geographical unit i. Our estimate for total suitable area across all units i ∈ Z where
P P P P (i)
Z is the set of all geographic units is then R = i∈Z Ri = i∈Z ai r c trc .

4. Application and Results

We used the complete model which combines statistical models for rooftop suitability,
plane size, tilt, and orientation to estimate the total amount of suitable rooftop area
in each of the 21 orientation classes for small buildings, on a zip code level across
the nation. Predicted values for all U.S. zip codes can be downloaded at [17]. Figure
5 shows the average estimated rooftop PV production per small building at the zip
code level. Broadly speaking, average small building production strongly correlates
with the solar resource; however, there exists significant local variation driven by
average household footprint and suitability [19]. For example, the simulated average
production in Florida is 12,100 kWh/year per small building (130% of the national
average) owing to an above-average solar resource, but it ranges from 5,300 kWh/year
to 30,100 kWh/year on a zip code level because of variation in suitability and building
footprint. Nationally, small building rooftops could accommodate an expected 731 GW

Pursuant to the DOE Public Access Plan, this document represents the authors' peer-reviewed, accepted
manuscript. The published version of the article is available from the relevant publisher.
of PV capacity and generate 926 TWh/year of PV energy on 4,920 km2 of suitable
rooftop space which equates to 25% the current U.S. electricity sales.

5. Conclusion

In this report we considered the technical potential of PV systems installed on exist-


ing, suitable roofs within the continental United States. Using a statistical modeling
approach, we made nationwide zip code level estimates of the number and charac-
teristics of small buildings, showing broad regional trends in both the suitability and
electric-generation potential of rooftops. Although only 26% of the total rooftop area
on small buildings is suitable for PV deployment, the sheer number of buildings in
this class gives them tremendous technical potential for energy generation.
It is generally understood that an individual household with adequate roof area
can generate greater than 100% of its annual electricity consumption with PV, yet
the variation in rooftop suitability and building characteristics makes it less obvious
whether that trend holds for groups of buildings. This analysis suggests that, in many
parts of the United States, zip-code-sized aggregations of households can collectively
generate enough electricity to offset their expected annual electricity consumption.
Beyond these conclusions about the potential of the rooftop photovoltaic resource,
we also extracted basic knowledge about the underlying processes that govern these
systems. That is, (1) the size of contiguous rooftop surfaces is well modeled by a
lognormal distribution, (2) the number of these planes is well modeled with an expo-
nential distribution, and (3) the fraction of buildings in a given region that are suitable
for photovoltaic installations can be effectively predicted using weighted multiple lin-
ear regression trained on demographic, economic, and land cover characteristics. This
paper provides the first investigation of technical potential for PV deployments on
existing rooftops that considers the size, shape, and orientation of rooftop surfaces as
they truly are “in the wild”.

Disclosure Statement

The authors declare that they have no conflict of interest.

Funding Details

This study was funded by the US Department of Energy Office of Energy Efficiency
and Renewable Energy (FY13 AOP SS13-FY13 AOP Systems Integration).

Data Availability

The data that support the findings of this study are openly available at
Data.gov, https://catalog.data.gov/dataset/u-s-pv-suitable-rooftop-resources-adac9.
The LiDAR data are available for visualization at http://maps.nrel.gov/pv-rooftop-
lidar.

Pursuant to the DOE Public Access Plan, this document represents the authors' peer-reviewed, accepted
manuscript. The published version of the article is available from the relevant publisher.
References

[1] Barbose, G., Darghouth, N., Weaver, S., and Wiser, R. (2014). Tracking the Sun VII: An
Historical Summary of the Installed Price of Photovoltaics in the United States from 1998
to 2013, Berkeley, CA: Lawrence Berkeley National Laboratory.
[2] Bright, J. and Burman, K. (2010). Portland Ecodistrict Solar Site Assessment, National
Renewable Energy Laboratory for Solar Cities America.
[3] Chaudhari, M., Frantzis, L., and Hoff, T.E. (2004). PV Grid Connected Market Potential
Under a Cost Breakthrough Scenario, 2004-117373. Navigant Consulting.
[4] Denholm, P. and Margolis, R. (2008). Supply Curves for Rooftop Solar PV-Generated
Electricity for the United States, NREL/TP-6A20-44073. Golden, CO: National Renew-
able Energy Laboratory.
[5] EIA (U.S. Energy Information Administration) (2009) Residential Energy Consumption
Survey (RECS), Washington, DC: U.S. Energy Information Administration.
[6] EIA (U.S. Energy Information Administration). 2012. Commercial Building Energy Con-
sumption Survey (CBECS), Washington, DC: U.S. Energy Information Administration.
[7] ESRI (2016). ArcGIS Desktop. Redlands, CA: Environmental Systems Research Institute.
[8] ESRI (2016). ArcGIS Desktop. Hillshade. 3D Analyst Toolbox. Redlands, CA:
Environmental Systems Research Institute. http://pro.arcgis.com/en/pro-app/
tool-reference/3d-analyst/hillshade.htm.
[9] Ferrari, S.L.P. and Cribari-Neto, F. (2004). Beta Regression for Modeling Rates and Pro-
portions, Journal of Applied Statistics, 31(7), 799-815.
[10] Frantzis, L., Graham, S., and Paidipati, J. (2007). California Rooftop Photovoltaic (PV)
Resource Assessment and Growth Potential by County, CEC-500-2007-048. Navigant Con-
sulting.
[11] Gagnon, P., Margolis, R., Melius, J., Phillips, C., and Elmore, R. (2016). Rooftop Solar
Photovoltaic Technical Potential in the United States: A Detailed Assessment, NREL/TP-
6A20-65298. Golden, CO. National Renewable Energy Laboratory.
[12] Land, C. E. (1971), Confidence intervals for linear functions of the normal mean and
variance, Annals of Mathematical Statistics, 42, 1187-1205.
[13] Margolis, R, Gagnon, P., Melius, J., Phillips, C., Elmore, R. (2017). Using GIS-based
methods and Lidar data to estimate rooftop solar technical potential in U.S. cities. Envi-
ronmental Research Letters, 12.
[14] Olsson, U. (2005), Confidence Intervals for the Mean of a Log-Normal Distribution, Jour-
nal of Statistics Education, Volume 13, Number 1.
[15] OSGeo: The Open Source Geospatial Foundation (2016) PostGIS: Spatial and Geographic
Objects for PostgreSQL.
[16] Paidipati, J., Frantzis, L., Sawyer, H., and Kurrasch, A. (2008). Rooftop Photovoltaics
Market Penetration Scenarios, work performed by Navigant Consulting, Inc. Burlington,
Massachusetts. NREL/SR-581-42306. Golden, CO: National Renewable Energy Labora-
tory
[17] Phillips, Caleb and Melius, Jenny (2016) U.S. PV-Suitable Rooftop Resources, National
Renewable Energy Laboratory.
[18] System Advisor Model (SAM) (2016) National Renewable Energy Laboratory. https:
//sam.nrel.gov/.
[19] Wilcox, S. M. (2012) National Solar Radiation Database 1991-2010 Update: User’s Man-
ual, National Renewable Energy Laboratory. NREL/TP-5500-54824.
[20] Zhang, X., Walker, R., Salisbury, M., Hromiko, R., and Schreiber, J. (2009). Creating a
Solar City: Determining the Potential of Solar Rooftop Systems in the City of Newark,
University of Delaware, Center for Energy and Environmental Policy.

10

Pursuant to the DOE Public Access Plan, this document represents the authors' peer-reviewed, accepted
manuscript. The published version of the article is available from the relevant publisher.

You might also like