Sunny Days for AI: How Deep Learning Changes Market for Solar-Powered Homes

by Tony Kontzer

Talk about solar-powered homes and the discussion quickly swerves to tax breaks, loans and other tools that help consumers cover equipment costs.

But price  isn’t the big hold-up, according to the founders of Oakland, Calif.-based startup PowerScout. It’s about marketing — and they’re using deep learning to fix it.

More specifically, they point to the industry’s reliance on an outdated door-to-door sales approach, in which salespeople leave fliers and seek one-on-one conversations with consumers.

“The cost of selling a solar system is more expensive than solar panels themselves,” said Kumar Dhuvur, co-founder and senior vice president of product for PowerScout, which is bringing intelligence and a powerful ecommerce platform to the solar market. “It’s almost how vacuum cleaners were sold in the 1960s.”

PowerScout’s goal is to change that model with an ecommerce site that leverages GPU-enabled deep learning to determine whether a household is likely to want solar, and what the feasibility and estimated value of a system would be. The company then reaches out to the most likely prospects, eliminating a lot of wasted sales and marketing costs.

PowerScout lidar
Sun city: PowerScout uses lidar, among other data, to zero in on homes likely to benefit from solar installations.

PowerScout’s Secret Sauce: Deep Learning

PowerScout’s deep learning-based analysis of a household’s likelihood to embrace solar, and its prospects for getting good solar production, is key. To train its deep learning convolutional neural network models, the company uses Amazon’s NVIDIA GPU-powered elastic compute cloud P2 instances, the CUDA parallel processing platform and the cuDNN deep neural network library.

So far, the company has trained two networks, both of which rely on analysis of satellite data: One determines whether a house already has solar panels; a second determines whether  vegetation is crowding the roof and could get in the way of an installation. The company plans to train more networks on GPUs, but is starting with the most obvious problems.

“Going out to identify how many homes have panels is impossible for a person to do, says Michael Ulin, a data scientist at PowerScout. “We’re always looking at improving our networks and improving our capabilities as our datasets grow. As we continue to learn this process, I think deep learning and NVIDIA GPUs are going to play a big role.”

Ulin says GPUs are vital to train PowerScout’s networks. “I would never think of training these models without GPUs,” he said.

Turning Over Every Rock

Kumar said that PowerScout’s approach of considering all the specific factors that contribute to a home’s solar-worthiness recognizes the unique potential of each home.

“The economics for each home are different,” said Kumar. “They face different directions, they shadow differently, they get more or less hours of direct sunshine. We are looking at everything pertaining to that address.”

PowerScout also presents manageable options for financing or purchasing a solar system for those consumers who want to move forward. It even matches consumers up with local installers it has certified.

Next on the company’s horizon: Branching into community solar — installations where multiple residents can take advantage. Further down the line, it plans to get into alternative energy products such as electric vehicles and power-storing batteries.

Said Kumar, “We’re just getting started.”