Building performance models (BPMs) have been used to simulate and analyze building performance during design. While extensive research efforts have made to improve the performance of BPMs, little attention has given to their robustness. Uncertainty is a crucial factor affecting the robustness of BPMs, in which such effect needs to be quantified through a suitable approach. The paper offers a robustness analysis framework for BPMs by using perturbation techniques to simulate uncertainty in input datasets. To investigate the efficacy of the framework, a generative adversarial network (GAN)-based framework was selected as a case study to analyze light switch usages in a single-occupancy office simulated using an immersive virtual environment (IVE). The robustness of the GAN was analyzed by comparing differences between a baseline (i.e., a BPM obtained from the GAN trained on a non-perturbed dataset) and BPMs obtained from the GAN trained on perturbed datasets. Overall, the robustness of the GAN significantly reduced when the training datasets were perturbed by using structured transformation techniques. The GAN remained relatively robust when the training datasets were perturbed by using an additive perturbation. Additionally, the sensitivity of the GAN involves different magnitudes corresponding to different levels of perturbed input datasets. The study suggests that the perturbation analysis is effective for investigating data uncertainty affecting the robustness of BPMs.

建筑性能模型 (BPM) 已用于在设计过程中模拟和分析建筑性能。虽然已经为提高 BPM 的性能做出了广泛的研究努力，但很少有人关注它们的稳健性。不确定性是影响 BPM 稳健性的关键因素，需要通过合适的方法量化这种影响。该论文通过使用扰动技术来模拟输入数据集中的不确定性，为 BPM 提供了稳健性分析框架。为了研究该框架的有效性，选择了基于生成对抗网络 (GAN) 的框架作为案例研究，以分析使用沉浸式虚拟环境 (IVE) 模拟的单人办公室中的灯开关使用情况。通过比较基线（即，从在非扰动数据集上训练的 GAN 获得的 BPM）和从在扰动数据集上训练的 GAN 获得的 BPM。总体而言，当使用结构化转换技术扰乱训练数据集时，GAN 的鲁棒性显着降低。当使用加性扰动对训练数据集进行扰动时，GAN 保持相对稳健。此外，GAN 的敏感性涉及对应于不同级别的扰动输入数据集的不同幅度。该研究表明，扰动分析对于调查影响 BPM 稳健性的数据不确定性是有效的。当使用结构化转换技术扰乱训练数据集时，GAN 的鲁棒性显着降低。当使用加性扰动对训练数据集进行扰动时，GAN 保持相对稳健。此外，GAN 的敏感性涉及对应于不同级别的扰动输入数据集的不同幅度。该研究表明，扰动分析对于调查影响 BPM 稳健性的数据不确定性是有效的。当使用结构化转换技术扰乱训练数据集时，GAN 的鲁棒性显着降低。当使用加性扰动对训练数据集进行扰动时，GAN 保持相对稳健。此外，GAN 的敏感性涉及对应于不同级别的扰动输入数据集的不同幅度。该研究表明，扰动分析对于调查影响 BPM 稳健性的数据不确定性是有效的。