Augmenting building performance predictions during design using generative adversarial networks and immersive virtual environments

Abstract

Existing building performance models (existing BPMs) often lack the capability for addressing human-building interactions in future buildings or buildings under design because they are mainly derived using data in existing buildings. The limitation may contribute to discrepancies between simulated and actual building performance. In a previous study, the authors discussed a framework using an artificial neural network (ANN)-based greedy algorithm which combines context-aware design-specific data obtained from immersive virtual environments (IVEs) with an existing BPM to enhance the simulations of human-building interactions in new designs. Although the framework has revealed the potential to improve simulations, it cannot determine the appropriate combination between context-aware design-specific data and the existing BPM.In this paper, the authors present a new computational framework (the GAN-based framework) to determine an appropriate combination based on a given performance target to achieve. Generative adversarial networks (GANs) are used to combine data of an existing BPM and context-aware design-specific data using a performance target as a guide to produce an augmented BPM. The effectiveness and the reliability of the GAN-based framework were validated using an IVE of a single occupancy office. Thirty people participated in an experiment on the simulation of artificial lighting switch uses using the IVE. Their light switch uses data under different work area illuminance were collected and analyzed. The building performance models (BPMs) proposed by Hunt and Da Silva were selected as the existing BPM and the performance target respectively. The data of each participant was used to generate an augmented BPM using the GAN-based framework and an updated BPM using the previous framework (i.e., ANN-based greedy algorithm framework). The thirty pairs of the augmented and updated BPMs were compared. Specifically, the errors measured between the updated BPMs and the performance target (E1) and the errors measured between the augmented BPMs and the performance target (E2) were analyzed using t-tests (α = 0.05). In 22 out of 30 cases, the performance of the augmented BPMs was significantly better than the updated BPMs, and in four cases, the performance of the two was similar. Only in four other cases, the performance of the updated BPMs was better. The results confirmed the efficacy of the framework. However, future research is needed to study the performance target and uncertainties associated with IVE experiments to better understand and control the reliability of the framework.