关键词: 深度学习, 数值模拟, 性能预测, 条件生成对抗网络, 高层居住区

Abstract: Modeling, simulation, analysis, integration, and optimization are frequently steps in the process of modifying the physical performance of home design schemes using computer numerical simulation software. This process has poor application value for several reasons: it takes a long time, decreases productivity, and has high hardware requirements. Despite these issues, architects require this numerical data. [Method] A case study based on two physical aspects (wind environment and sunshine environment) of residential areas was carried out using computer deep learning. A performance simulation image prediction model based on conditional generative adversarial network (CGAN) was built by learning the overall layout plane of residential areas under different height distribution characteristics as well as the corresponding sunshine simulation and wind environment simulation images. Construction of the performance simulation image prediction model covered training dataset establishment, physical performance training mode, Tensor Flow tools, and CGAN data structure. Specifically, the training dataset can apply multiple methods. Physical performance training mode can be divided by type based on physical model and the type of data. It defines the generator network, discriminator network, and loss function by using Tensor Flow tools. The data structure of CGAN includes determination of the generator network, discriminator network, and loss function. The objective image quality assessment algorithm (SSIM) based on structural similarity was used to analyze and validate the predictions of the trained model. The experimental findings prove CGAN efficacy in image generation. Therefore, CGAN can realize fast prediction and generation of performance simulation images with limited errors and has considerable practical value. In addition, the performance simulation images produced by CGAN can facilitate adjustment and optimization of the overall layout. Based on the open source programming platform, the performance prediction model was trained and built using CGAN. The parameters and loss function were adjusted continuously by producing the training dataset and building the generator network and discriminator network for the performance prediction model. The final performance prediction model can predict and generate the corresponding performance simulation images quickly and accurately after inputing the overall layout with grey value information. Based on predicted and generated performance simulation images, architects can adjust and optimize, increasing working efficiency. This has strong practicability in the preliminary design stage of residential areas. The design method in this experiment has some limitations. Firstly, the previous layout data for training of the neural network was collected from the typical point-plate mixed residential areas in Hangzhou. Hence, the trained neural network is only applicable to point-plate mixed residential areas and is inapplicable to other layouts. Secondly, the layout scheme used as the training dataset is based on fixed plots. The final results can provide design references to high-rise building distributions in any plots, but they are not completely applicable. This will be addressed further in future studies.

Key words: deep learning, numerical simulation, performance prediction, conditional generation adversarial network, high-rise residential area