DeepSolar-3M: An AI-Enabled Solar PV Database Documenting 3 Million Systems Across the US (Papers Track)
Rajanie Prabha (Stanford University); Zhecheng Wang (Stanford University); Chad Zanocco (Stanford University); June Flora (Stanford University); Ram Rajagopal (Stanford University)
Abstract
The widespread deployment of distributed energy resources (DERs), especially solar photovoltaic (PV) systems, is essential for attaining a sustainable energy future. Uneven solar adoption slows decarbonization and entrenches energy poverty in vulnerable communities. ML-driven spatial analytics can pinpoint these gaps to prioritize policy interventions. Granular spatial mapping of all installed PV systems provides comprehensive data to facilitate a more equitable and efficient transition to sustainable energy sources. Previously, DeepSolar provided the most comprehensive nationwide solar PV dataset in the United States; however, it only extends up to mid-2017. This paper introduces a novel pipeline leveraging vision transformer models to detect rooftop-mounted solar PV systems at the building level, extending coverage through 2022. Our findings indicate that rooftop PV systems in the U.S. have doubled over the past five years, totaling approximately 2.95 million systems, with increased adoption observed across all states. The final dataset, which we are making publicly accessible, serves as an invaluable resource for policymakers, developers, researchers, and utilities dedicated to advancing equitable decarbonization efforts nationwide.