‘DeepRoof’ AI targets cities’ solar hot spots

Source: By David Iaconangelo, E&E News reporter • Posted: Tuesday, August 13, 2019

Computer scientists from the University of Massachusetts, Amherst, say they’ve invented a tool that can zero in on rooftops well-suited for solar panels in nearly any part of the country.

Researchers developed the software tool, known as DeepRoof, with homeowners in mind, said Stephen Lee, associate professor in the University of Pittsburgh’s computer science department and lead author of a paper about the model.

But as city and state authorities prepare to meet emissions reduction goals and transition to renewable power sources, policymakers and major solar installers could find it useful, as well, he added.

“They could see how effective solar would be from a city-scale level,” said Lee. “Or for targeted solar initiatives — is this part of the city good for solar?

“We had the private sector in mind when we designed it, but I see broader applications elsewhere,” he added.

Lee’s team presented the paper this week at a conference this week hosted by the Association for Computing Machinery in Anchorage, Alaska.

Their effort marks the latest attempt to automate the work of evaluating a particular rooftop’s solar feasibility, a task that has traditionally required professional consultants to carry out on-site.

Existing tools, like Google’s Project Sunroof and the Massachusetts Institute of Technology-backed Mapdwell, rely on lidar technology — a sort of laser that measures distances between objects — to remotely create a 3D image of a given rooftop.

The National Renewable Energy Laboratory (NREL) has used lidar for its own projections of solar’s technical potential.

But lidar images only cover certain parts of the country: about 128 metropolitan areas, accounting for 40% of the U.S. population, according to NREL.

By contrast, DeepRoof uses machine learning to translate 2D satellite images, readily available by Google or Bing Maps, into 3D mockups of individual rooftops.

Based on those images, the tool assesses key features about the shape and orientation of the roof, then compares those figures with public records on the region’s solar irradiance, specific building heights and number of floors.

DeepRoof has a harder time estimating the installation area for smaller rooftops, where certain features can be harder to identify from satellite imagery, according to the paper. But its scores start to grow more reliable as roof sizes increase: with rooftops between 2,000 and 3,000 square feet, DeepRoof and Google’s Project Sunroof deliver nearly identical estimates, researchers said.

Lee said the next project would be to integrate the tool with another model that could calculate the cost of adding solar panels to each site. But DeepRoof is a “first step” toward piquing homeowners’ interest in getting solar installed, he said.