当前位置: X-MOL 学术Energy Build. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Developing a dual-modal surrogate model training framework for building performance prediction in early design stage
Energy and Buildings ( IF 6.6 ) Pub Date : 2025-01-14 , DOI: 10.1016/j.enbuild.2025.115307
Zhen Han, Gang Liu, Lihua Zhang, Xiaoqian Li, Ye Yuan

In the early stages of architectural design, machine learning-based surrogate models exhibit significant advantages in rapidly evaluating building performance compared to traditional simulation software. Currently, most building performance prediction surrogate models rely on a single data type as input for performance prediction, which limits the model's generalizability. This study proposes a novel training framework based on dual networks: a predictor and an inverse predictor, each incorporating corresponding modality fusion strategies. This allows the prediction model to simultaneously accept both image and vector modalities as input. Images are used to describe architectural layout features, while vectors represent other information such as meteorological conditions, window-to-wall ratio, and thermal transmittance, etc. Compared to traditional models with a single predictor, the inverse predictor introduced in this study allows the prediction model to focus more on the modality with lower data volume (vector modality) without causing overfitting, achieving lightweight and high-precision predictions. Through case studies on residential building energy consumption and daylighting performance surrogate models, the proposed dual-modal prediction method demonstrated an R2 of 0.9250 on the training set and 0.9040 on the test set, with MSEs of 0.0011 and 0.0014, respectively. These results validate the model's effectiveness and accuracy, offering a new methodological framework for dual-modal architectural performance prediction.

中文翻译:


开发双模态代理模型训练框架,用于在早期设计阶段进行建筑性能预测



在建筑设计的早期阶段,与传统仿真软件相比,基于机器学习的代理模型在快速评估建筑性能方面表现出显著优势。目前,大多数建筑性能预测代理模型都依赖于单个数据类型作为性能预测的输入,这限制了模型的泛化性。本研究提出了一种基于双网络的新型训练框架:预测器和逆向预测器,每个框架都包含相应的模态融合策略。这允许预测模型同时接受图像和矢量模态作为输入。图像用于描述建筑布局特征,而矢量表示其他信息,例如气象条件、窗墙比和热透射率等。与具有单个预测器的传统模型相比,本研究中引入的逆向预测器使预测模型能够更多地关注数据量较小的模态(向量模态),而不会造成过拟合,实现轻量级和高精度的预测。通过对住宅建筑能耗和采光性能代理模型的案例研究,所提出的双模态预测方法表明,训练集上的R2为0.9250,在测试集上的R2为0.9040,MSE分别为0.0011和0.0014。这些结果验证了模型的有效性和准确性,为双模态架构性能预测提供了新的方法框架。
更新日期:2025-01-14
down
wechat
bug