SolarGAN: Synthetic Annual Solar Irradiance Time Series on Urban Building Facades via Deep Generative Networks

Abstract

A generative model using simple fisheye images as inputs to predict solar irradiance time series on building facades in urban environments• Fisheye images as categorical shading masks are captured from opensource LOD 1 urban geometry, but in principle may also stem from real photographs• Model architecture allows for generation of stochastic ensembles of annual hourly time series, as well as good consistency to the deterministic data• Developed a network architecture combining Variational Autoencoder (VAE) and Generative Adversarial Networks (GAN)• Developed model addresses inherent uncertainty of intermittent solar energy and allows parametrically altering urban environmental features

Stochastic Solar Irradiance from Deep Generative Networks and their Application in BIPV Design

“…Finally, we exhibit its application potential on district multi-energy systems (d-MES) and solar-driven geometry design. Due to space limitations, we provide details and specifics of the model architecture separately in a preprint [18] while focusing on the application potential here.…”

“…It fuses the VAE-extracted urban context low-dimensional representations together with other conditions for each test point and generates an ensemble of annual solar irradiance time-series for each one, which are consistent with the given conditions while exhibiting stochasticity in terms of weather patterns. The details and specifics of the SolarGAN model architecture are provided in another preprint [18].…”