Publikation im Journal "Applied Energy"

27. Dezember 2021

Unsere Studie Large-Scale Detection of Rooftop-Mounted Photovoltaic Systems in 3D wurde in der Fachzeitschrift Applied Energy (forthcoming) aufgenommen. Das Paper wurde in Kollaboration mit Forschern der LMU München (Marie-Louise Arlt) und der Stanford University (Kevin Mayer, Zhecheng Wang und Ram Rajagopal) erarbeitet.

Autoren: Kevin Mayer, Benjamin Rausch, Marie-Louise Arlt, Gunther Gust, Zhecheng Wang, Dirk Neumann, Ram Rajagopal

Abstract:

While photovoltaic (PV) systems are being installed at an unprecedented rate, it is challenging to keep track of them due to their decentralized character and large number. In this paper, we present the 3D-PV-Locator for large-scale detection of roof-mounted PV systems in three dimensions (3D). The 3D-PV-Locator combines information extracted from aerial images and 3D building data by means of deep neural networks for image classification and segmentation, as well as 3D spatial data processing techniques. It thereby extends existing approaches for the automated detection of PV systems from aerial images by also providing their azimuth and tilt angles. We evaluate the 3D-PV-Locator using a large dataset gathered from the o_cial German PV registry in a real-world study with more than one million buildings. In terms of azimuth and tilt angles, our evaluation shows that the 3D-PV-Locator and the o_cial registry coincide for about two thirds of the observations and are within neighboring classes for 84 and 99 percent of the observations, respectively. In terms of detected PV system capacity, we show that the 3D-PV-Locator clearly outperforms existing approaches. It performs particularly well for the groups of small and medium-sized PV systems (3.6-33.1 percent error reduction) and PV systems tilted beyond 40° (25.6-38.1 percent error reduction). The 3D PV system data generated by the 3D-PV-Locator can inform several practical applications, such as improved forecasting of solar generation, the optimized planning and operation of distribution networks, improved integration of electric vehicles, and others. All datasets and pre-trained models associated with this paper are available online.