DemAEndGAN: Cross-modal hourly building energy demand prediction with autoencoders and generative adversarial networks

While the existing literature shows numerous Machine Learning (ML) based approaches for deterministic aggregate building energy performance prediction, stochastic time-resolved models are still being explored. However, such features of a model are required to address contemporary challenges, such as the integration of intermittent renewable energy or occupant-driven controls. This study therefore proposes a novel data-driven generative ML architecture, ”demAEndGAN”, for the prediction of hourly stochastic building energy demand profiles. Unlike other time-resolved data-driven building energy models, which are functions of explicit, numerical descriptions of building characteristics, our approach uses isometric images of a building as predictors. It is made possible by a novel cross-modal architecture combining a Variational Autoencoder (VAE) and a Conditional Generative Adversarial Network (cGAN). Trained on a large dataset of 200k building energy simulation samples, “demAEndGAN” is able to generate accurate annual hourly time series for ideal cooling and heating demand of various climates, with a median Normalised Root Mean Square Error (NRMSE) of 6–9%. Furthermore, with the variance-based encodings in its VAE module and random noise injections in its GAN module, “demAEndGAN” also demonstrates the ability to generate stochastic demand predictions, as well as generate parametric alterations of the original input isometry with corresponding physically coherent energy demand predictions. In conclusion, this novel approach has the potential to be reliably utilised in interactive early-stage decision making to (visually) traverse the design space, as well as to conduct large-scale scenario-based robust urban energy planning.