On a bright morning, graduate student Jeremy Klotz and professor Shree Nayar walked through upper Manhattan with a tall tripod and a camera that takes 360-degree images. Their route took them to bike docking stations, which use solar energy to power their kiosks, docking mechanisms, wireless communication, and even E-bike recharging in recent installations. At each docking station, the researchers raised the camera above the panel, snapped a spherical picture, and sent it to Klotz’s laptop.
Seconds later, the team’s computer vision program told them something remarkable: how much energy that panel would generate in a year — and how much it could generate if it were pointed at the optimal angle.
As it turns out, the solar panels powering the bike docking stations — and likely many solar panels across New York City and other urban destinations — may be leaving significant energy untapped simply because they are not at their best orientation.
By reorienting the panel, Klotz and Nayar, T.C. Chang Professor of Computer Science at Columbia Engineering, estimate it could harvest up to 30% more energy over the course of a year. They found the same story at several other bikeshare docking stations in the neighborhood.
“Our research result makes it possible to make important decisions such as where to place solar panels, how to angle them, and what the return on investment of an installation will be in the long run,” Nayar said. This is a surprising technical result because a panel installed in a dense urban setting only sees a part of the sky and must contend with cast shadows and reflections from buildings and other structures around it.
Using a single image, the technology predicts how shadows, reflections, and likely weather patterns will affect the amount of solar energy that reaches any location in any city, automatically accounting for the sun’s daily and annual movement through the sky. The technique, which operates on off-the-shelf hardware, is described in a research paper published April 16 in Solar Energy.
Solar has become one of the cheapest sources of energy. The price of solar panels has fallen so steeply that the expense of installing a panel — such as permits, labor, and site assessment — costs more than the panel itself. Knowing in advance how much energy a given surface will actually produce is key to determining whether an installation makes financial sense.
Solar farms optimize energy production by placing rows of solar panels in open fields, maximizing the amount of direct sunlight that reaches the panels. Some have motors that track the sun through the day and across the year.
The situation in cities is different. Urban solar panels help stabilize the power grid and lower energy prices by generating energy where demand is high and growing. But there’s a tradeoff: objects like buildings, utility poles, water towers, and signs block direct sunlight, limiting the energy that a panel can produce. In the solar industry, these locations are called urban canyons. The existing methods for predicting how much energy a solar panel in an urban canyon will produce are time-consuming, expensive, and notoriously inaccurate at predicting the energy it will receive from three sources: direct sunlight, diffuse light scattered across a partially visible sky, and light reflected off surrounding buildings.
“In solar harvesting, small objects that are close to a panel can have as significant an impact as really large objects that are far away,” Nayar said. “On a roof, for example, relatively small objects like HVAC units or parapets don’t always appear in 3D models of cities, but they have enormous impacts on the light a panel receives.”
Diffuse and reflected light further compounds the problem. Standard methods produce large errors by failing to account for how widely the brightness of a section of sky visible to a panel varies throughout the day and year, and by ignoring light reflected off surrounding buildings. The Nayar lab found these reflections account for roughly 12% of a panel's total annual energy on average.
“One would imagine that reflected light makes a negligible contribution, but actually it turns out to be significant,” Nayar said. “If a building in the panel’s field of view is being directly lit by the sun while the panel is in shadow, that reflected light will account for most of the energy the panel receives.”
The basic insight behind the new method is that a photograph taken from a panel's location encodes an enormous amount of information. Shadows in the image reveal the direction of the sun, and straight lines in the architecture indicate the direction of gravity. This information is used to find the image's orientation with respect to the planet. A segmentation algorithm reveals the portion of the sky that is visible to the panel, and the appearance of buildings around the panel provides cues related to the spatial layout of the buildings and the materials they are made of.
From there, the system forecasts how the sun will move through that visible sky across the day and year, estimates how much light will bounce off surrounding buildings onto the panel, and draws on historical weather data to extend the forecast across cloudy and overcast conditions, not just clear days.
The computation happens almost instantaneously on a laptop.The method works not just for panels on rooftops and pole-mounted panels at ground level, but also for vertical walls.
“This technology is inexpensive and portable, making it usable by both homeowners as well as large companies during the planning stage of a solar project to maximize the return on their investment," Nayar said. "Right now, that process is slow, expensive, and approximate at best."
For the first time in decades, electricity demand is growing faster than electrical grids can keep up. Data centers, heat pumps, and electric vehicle charging are forcing utility companies to make expensive infrastructure investments that ultimately show up in consumers' bills.
Vijay Modi , a professor of mechanical engineering and of earth and environmental engineering at Columbia who studies energy systems in cities, sees Klotz and Nayar’s technology as a means of increasing the amount of solar panels that are installed.
“If you don't have a tool to assess, you can't deploy at scale," Modi said. "Quantitative time-series solar potential that the technology provides allows one to assess value to both the end-use and the grid- key to making investment decisions."
Modi's own research has documented one underexplored source of that potential. The facades of tall buildings, he argues, represent a significant opportunity because they have far larger surface areas than rooftops. Most tall buildings receive more direct sunlight than the roof. Using Klotz and Nayar's method, his team measured the east-facing wall of a building on Columbia's campus and found solar radiation levels high enough to meaningfully offset both the building's energy use and the peak electricity demand. It was, he said, a result that could be obtained in hours, and without the tool would have taken a month with expensive instrumentation to carry out.
"Increasingly, the challenge in cities is that as loads grow, the utility needs solutions that can offset distribution network upgrades — the prime driver for high tariffs for consumers," Modi said. "We have not recognized that tall buildings can be contributors to a solution through self-generation and storage in the future, and a source of peak loads. Practically every building I can see from my office window could be producing solar energy on its walls."
For Nayar, the technology points to a future in which cities generate far more of the energy they rely on.
"At the price panels are today, virtually any surface in the world has the potential to be an energy harvester," he said. "The question is which ones are actually worth it."
Klotz and Nayar have filed for a patent on the new technology.
Solar Energy
Forecasting solar energy using a single image
16-Apr-2026