Zhang 2022 IOP Conf Ser Earth Environ Sci 1078 012040oCov
This 2022 conference paper (Zhang, Waibel, Schlüter) introduces SolarGAN, a data-driven surrogate that combines image VAE and conditional time-series GANs to generate high-resolution, stochastic ensembles of annual hourly solar irradiance for building envelopes using simple fisheye categorical mask images (from LOD1 urban geometry) plus auxiliary inputs (weather stats, location, height, normal). The model is trained on simulated ground truth (ClimateStudio) for five European cities and validated statistically (JSD, MAE) and in two applications: resilient district multi-energy system (d‑MES) design and parametric solar-driven geometry design. Authors report substantial speedups versus full simulation and discuss current limits (dataset coverage, long-horizon fidelity, generalization).
Questions this document answers
10.1088/1755-1315/1078/1/012040
September 2022
Geneva, Milan, Paris, Berlin, Zurich
5000 test points; 15,000 images (3 WWR levels); 75,000 annual time-series (5 locations); train:test = 4:1
40 stochastic time-series per test point