Solar Energy Adoption by County in the US

Or: Does Google Have All the Data I’ll ever Need?

He says you should have solar panels.

Introduction

What with all the hubbub about climate change, potential terrorist attacks on the power grid, climate change-related severe weather destabilizing said grid, and the continually decreasing cost of solar panel installation, one could be forgiven for thinking that everyone in the know has rushed out and had solar panels installed on their roof.

For those who want to explore solar as an option, who else to turn to but Google? Their Project Sunroof can tell you how much energy-generating potential your roof has, guide you to incentives for solar installation, and give you information about the solar suitability and current solar installations for any neighborhood, city, county, or state where they have sufficient data.

As someone deeply concerned about climate change, but also afflicted with the dreaded FOMO (since my heavily shaded house has no solar panel potential), I was curious about which demographic factors drive solar energy adoption and how quickly it is being adopted.

The variables I decided to analyze, and why I chose them, are:
- Per-capita income: is solar widespread, or a high-income luxury?
- Population density: is solar more prevalent in urban or rural areas?
- Homeowner status: renters probably can’t put panels on their roofs, at least not of their own volition.

I also looked at areas with the largest and smallest growth (based on change in percentage of qualified roofs with installations) for 2021 versus 2017. This might give insight into which regions could be targeted for increased marketing of solar, which might be more receptive to political initiatives promoting renewable energy, or to where a solar-FOMO afflicted person might want to relocate.

I picked county-level data as a compromise between state- and city-level.

The primary data source was Google Project Sunroof . 2021 data were obtained directly from the site, while 2017 data were obtained from Kaggle.

US census data were used to obtain per-capita income, population density, and homeowner status. Income and population came from Kaggle; land area from 2010 census data, and homeowner status from the 2019 community survey.

The data

I kept only counties where Google had data on at least 50% of homes in both 2017 and 2021. When combined with census data, this left 481 counties, representing 49 states, out of 3142. This is something often glossed over when data analysis is discussed: 100% complete data is almost never available. What you’re reading about is almost always based on some smaller sample that may or may not represent actual reality.

The results

Only about 0.2% of the roofs Google surveyed were estimated to have solar panels. This was the same for 2017 and 2021.

Question 1: Is per-capita income associated with increased solar adoption?

The plot below doesn’t show much of a relationship between income and solar adoption. There might be a slight tendency for more installations at higher per-capita income, but there aren’t many data points to base such a conclusion on.

Question 2: Is population density associated with increased solar adoption?

Population density doesn’t show much association with solar panel adoption either. The graph below is plotted on what’s called a logarithmic scale to spread out the data at the low end more evenly, but there’s still not an obvious relationship.

Question 3: Is home ownership associated with increased solar adoption?

While it looks like areas with low home ownership (< 30%) tend to have less solar, there aren’t many counties in that range, and most of the data don’t show much of a relationship.

Question 4: Which counties showed the most and least growth in solar adoption?

The Project Sunroof data didn’t show much change from 2017 to 2021 in the percentage of eligible roofs with solar installations. Actually, on average it showed a tiny (-0.02%) decrease. Of the top 10 counties, only two had an increase, and the maximum was 0.12%. The bottom 10 counties decreased by -0.7% to -1.5%. Kentucky had the most counties (2) in the top 10, while Florida had the most (4) in the bottom 10.

It’s possible the decrease could be because Project Sunroof surveyed more land or houses in 2021 than in 2017. But, the data show coverage was actually a little lower (5.0% less).

More houses might be being deemed solar-qualified by Google’s algorithm. In fact, 10.7% more were judged to be solar-qualified in 2021. However, this didn’t seem to have a very strong relationship with the percentage of houses with solar.

So what does and doesn’t this mean?

There wasn’t a clear relationship between the variables I analyzed (per-capita income, population density, and home ownership) and solar adoption, at least using the Project Sunroof data.

This might be because there are so few solar installations identified (about 0.2% of roofs). When something is this rare, it’s hard to figure out what it’s influenced by without tons of data. It is also possible that state or local incentives (like tax credits or net metering), zoning or homeowner association restrictions (which would trip me up even if my neighborhood’s trees didn’t), and availability of solar installers could have much more influence than the variables I analyzed. I couldn’t analyze these other variables with the dataset I used.

Also, the data (with some manipulations to account for 32 counties with missing home ownership values) cover 481 counties in 49 states, out of 3142 counties in the 2017 Census data. This might not be enough to get an accurate estimate of the true adoption of solar energy.

Another thing to consider is that Project Sunroof might not be accurately estimating the number of installations. They estimate installs with a machine-learning algorithm trained on human-labeled images of roofs. But, they don’t tell us how accurate the algorithm is or if they tested it against something like how many solar panels were installed or sold. Data that’s been more fully validated might paint a different picture.

The “growth” data I calculated showed no change or a slight decrease in solar adoption. A true decrease, as in people removing solar panels, isn’t likely, as good-quality panels last 25–30 years. Solar energy industry data also indicate a steady rise in residential installations between 2017 and 2021. But, this only represents about 2000–3000 homes/year for the entire country, a very small percentage of all homes. Additionally, the growth data from Google depend on their machine-learning algorithm, as discussed above. We don’t know its accuracy, so we can’t really trust its growth estimates.

In summary, this analysis couldn’t identify a relationship between demographic variables and solar adoption for selected US counties. There might be variables that weren’t in the dataset that are important. Estimating growth in solar adoption using the Project Sunroof data showed almost no increase in installations as a percentage of roofs identified as eligible. This might be because the model can’t accurately identify solar installations, or because the sample isn’t representative. So, to answer the subtitle’s question: No, Google (and most data sources) don’t have all the information you need.