Stanford University has rolled out "DeepSolar" software and open platform, capable of locating 1.47 million PV power devices in the U.S. via machine learning and satellite images, which can help people have a grip on the local deployment of PV power systems.
Amid the green-energy craze, output value of the PV power device installation industry in the U.S. shot up to US$210 million in 2017, up from US$42 million 10 years ago. There, however, still lack sufficient related statistics, such as distribution of PV power systems, installation motive, hot installation spots, and sunshine conditions, which has hampered the works of policy makers, scientists, and PV power developers.
"DeepSolar" will facilitate access to such vital information, helping developers formulate feasible development plans, utilities attain balanced grid networks and set up feed lines, and municipal governments carry out urban development planning.
The Stanford University research team had computer recognize 370,000 satellite images sized 30 x 30 meter and then determined whether they contain solar panels, via an algorithm which divides the images into multiple small images for further processing and classification by an in-depth neural network before identification of the containment of solar panels and formation of heat map.
Jiafan Yu, a doctorate-degree student for electrical engineering, pointed out that DeepSolar has attained 93.1% precision and 88.5% recall in the search for residential PV power systems, compared with 93.7% and 90.5%, respectively, for non-residential systems.
The research team has had DeepSolar identify 1 billion satellite images, thereby locating 1.47 PV power devices, much higher than original estimate of 1.02 million, including residential and office-building rooftop systems and large-scale PV power stations, in a span of one month. The team will combine the figure with census result and other data in analyzing the motive of people in installing PV power devices.
The team found that areas with highest density of PV power devices are those with population density reaching 1,000 persons/square mile, especially places with average annual household income topping US$150,000. The study also shows that combination of the data on the sites of PV power devices and geographic data can exhibit the minimum sunshine amount for the setup of such devices.
Arun Majumdar, associate professor for civil and environmental engineering at Stanford U., pointed out that the team will publicize other data produced by DeepSolar successively and extend the scope of its research to foreign areas.
All related data have been available on DeepSolar website and the study has been published in "Joule."
- Stanford’s DeepSolar AI maps solar installations right across the US
- Stanford scientists locate nearly all U.S. solar panels by applying machine learning to a billion satellite images
- This AI Just Mapped Every Solar Panel in the United States
(First photo courtesy of Stanford U.; written by Daisy Chuang)
note: recall refers to sensitivity, or the percentage of correct judgment in the samples with positive answers.