一种高层建筑群排布生成与推荐算法

An Algorithm for Generation and Recommendation of Highrise Building Group Layout

针对高层建筑物排布在日照等多个指标的自动优化问题,提出了基于强化学习的高层住宅自动布局设计方法,根据建筑设计中的建筑设计防火,城市居住区规划设计等规范要求设计奖励函数。通过各建筑单体与环境的不断交互、训练学习,输出多个满足设定约束的布局方案。然后对方案中的建筑群进行基于傅里叶描述子的模式表示和自动聚类,提供去冗余后的多元典型设计模式,为建筑设计师提供设计参考。最后,在北京某地块的真实环境中验证了方法的可行性和有效性。

With consideration for the norms and constraints of highrise building layouts regarding fireproof building design and residential district planning, an automatic layout method for high-rise residential buildings based on reinforcement learning and unsupervised learning was proposed. It provides references for automatic generation of building layouts and pattern clustering of building clusters for architects and designers.Firstly, building cluster layout was designed by using a deep deterministic policy gradient algorithm based on reinforcement learning.A reward function with considerations for design norms of land use, fire separation distance, and sunshine duration was designed. Specifically, the bonus value of land use redline was used as the overlapping area between areas enclosed by four vertexes of the current building and the involved land area. The bonus value of fire separation distance was designed as outward rounded rectangles by centering the involved building to make distances from each point to the building monomers agree with specific distance scope. This was used as the fire prevention region. The overlapping area between the fire prevention region of the building under planning and the fire prevention region of other buildings was calculated. A bonus value of sunshine duration was designed and calculated from each frame. Meanwhile,the number of sunshine test points which fail to meet the sunshine duration as well as the calculation of sunshine duration at the sunshine test points of each building were provided. This model inputs the initial layout given by architects and each building monomer was moved according to the given order. The building was fixed after moving for several steps, thus completing the layout of all buildings in order. Finally, the layout generated by the model was judged. If the design scheme meets all constraints, it is a reasonable scheme. Several layout schemes that meet the presetting constraints were output through continuous interaction, training, and learning between building monomers and environments.Next, the generated reasonable layouts were expressed in modes. The buildings were extracted as points. Centroids of different individual buildings were chosen to form the pattern shape of building cluster. Pattern expressions of the generated reasonable layouts were realized using Fourier descriptors.With consideration for the rotational invariance of Fourier descriptors, the directional characteristics of building cluster were added. The minimum external rectangle expressed in each layout pattern was chosen and the included angle between the long-axis of the rectangle and positive direction of x-axis were used as the directional characteristics of building clusters.The seven-dimensional characteristics were normalized and the generated reasonable layouts were clustered using a K-means clustering algorithm.Different clusters with similar building patterns were formed and a multielement typical design mode after elimination of redundancy was given,which provided references to architects.Finally, the feasibility and effectiveness of the proposed method were verified in a real environment in a block in Beijing. A total of 87 groups of schemes were output over 80,000 rounds. Moreover, the generated schemes were divided into three types to be used as design references by architects.Compared with traditional automatic design processes, the proposed method doesn’t need a layout of samples and can provide several layout schemes meeting the preset constraints for a block.