DeepSolar++

Solar photovoltaic (PV) systems are being deployed at a rapid yet non-uniform pace. To explain this heterogeneity across space and time, in the DeepSolar++ project, we applied computer vision to historical satellite and aerial images to uncover when solar PVs have been installed on top of the DeepSolar project that investigated the problem of where. We, therefore, constructed a nationwide spatiotemporal dataset of PV deployment in the United States. We analyzed the data using a technology diffusion model and found that low-income communities are not only delayed in their PV adoption onset but also saturate more quickly at lower levels. We also found that certain types of incentives are associated with a high saturated adoption level in low-income communities, whereas other types are not. The computer vision model we developed can be scaled to any location on Earth where historical aerial or satellite images are available at a sufficient resolution. Additionally, we make our dataset publicly available as a resource for researchers, policymakers, and other stakeholders to understand PV adoption dynamics and customize policy design.

Publication:

• Zhecheng Wang, Marie-Louise Arlt, Chad Zanocco, Arun Majumdar, and Ram Rajagopal (2022). DeepSolar++: Understanding Residential Solar Adoption Trajectories with Computer Vision and Technology Diffusion Models. Joule. (link) (project website) (code) (Stanford News)

• Zhengcheng Wang*, Zhecheng Wang*, Arun Majumdar, and Ram Rajagopal (2019). Identify Solar Panels in Low Resolution Satellite Imagery with Siamese Architecture and Cross-Correlation. NeurIPS Tackling Climate Change with Machine Learning Workshop. (* Equal contribution) (link)