生成建築配置圖中使用機器學習整合採光與能耗的方法

Abstract

敷地計畫在建築平面設計中是重要的階段，需要考慮包括考慮場地環境和永續議題等。過去研究雖然已經探討了使用基因算法和機器學習的自動平面圖生成，但往往忽略了基地條件，導致方案不能回應實際設計。此外儘管基因演算法很常用於平面生成，但其計算成本高且費時。本文處理前期建築配置的自動生成，主要強調將基地條件的納入初期建築生成設計，尤其著重物理環境造成的影響，而這個階段的策略也將對後續設計產生重大影響。 本文提出了一種新的自動生成配置圖的流程，將法規、採光與能耗等因素納入考慮，並且以基地條件作為約束與設計的新可能性。作者提出最佳化前期平面圖生成的流程，將參數化和機器學習方法整合，而法規、日照和能源消耗等面向作為設計目標，提供設計師更為永續的解決方案。作者使用參數化模型生成平面圖，並且嘗試不同的生成式機器學習模型。接著使用pix2pix和Resnet-50評估採光與能源消耗結果，最後採用本地搜索演算法來建議最佳的設計方案。 主要貢獻包括： 1.在自動建築平面生成流程中，整合基地條件。 2.通過增加設計要素和擴展數據集來增強評估階段，以提高精度和適用性。 3.利用機器學習生成的結果以及採光和能源消耗的評估數值來尋找最佳化的平面圖。 這篇論文提供自動化建築平面設計更完整與有效的永續設計方法，以助於建築師和研究人員更快速產生質量更高的初步設計方案。

Site planning stands as a pivotal stage within architectural floor plan design, including considerations of site contexts and sustainable issues. While previous research has explored floor plan generation using genetic algorithms and machine learning, it often neglected site context, resulting in undesired solutions. Moreover, though genetic algorithms are common, they can be computationally intensive. This thesis ad-dresses machine learning-based early-stage architectural plan generation, emphasizing the inclusion of site conditions, particular for influences from sunlight. The goal is to find a sustainable suggestion for footprint, which will greatly impact subsequent design. This thesis presents a novel approach to automated site planning that incorporates regulations, daylight and energy considerations. The author introduces a pipeline for optimal footprint generation through parametric and machine learning methods, ad-dressing the lack of integration between design phases and offering a more site-responsive and eco-friendly design approach. The author generates footprint with parametric models, experiments generative machine learning models, and evaluates the results using pix2pix and Resnet-50. A local search is then applied to suggest optimal solutions. Key contributions include: 1.Pioneering the integration of site conditions into the automatic architecture plan generation pipeline. 2.Enhancing the evaluation pipeline with additional design factors and a larger dataset for improved accuracy and applicability. 3.Utilizing machine learning outcomes to optimize plans while providing evaluation feedback on daylight and energy consumption.