基於深度學習模型ANN、3D-CNN與3D-Unet優化環境性能驅動生成式設計

由於科技的不斷進步，使得人工智慧得以蓬勃的發展，現今社會的運作已經部分委由人工智慧來執行，藉由人工智慧的輔助，我們能夠大幅減少人力成本以及增加任務執行的效率。由於人工智慧工具的應用門檻已經被降低，使得這項工具能夠擔任的工作越來越普遍，進而使得各行各業也都開始考慮如何使用人工智慧結合各自的領域專業。建築的組成非常複雜，所考量的因素非常廣泛，並且在成本上也是非常昂貴，因此需要有建築相關背景的研究者來研究建築與人工智慧的結合方式。本研究認為建築對於周遭環境與社會的影響不容小覷，且尚有許多棘手的問題等待被解決，而透過傳統人力是不足去解決這些論題的，因此本論文嘗試使用人工智慧去解決建築方面與建築環境的相關問題。建築環境是建築師非常關注的一項議題，可以藉由環境性能模擬去評估所設計的建築量體或者設備系統是否能夠達成使用標準。然而今日的環境模擬系統需要耗費非常長的時間與電腦資源去計算環境數值，也因此在現況中，建築的環境評估被置放於設計後期階段實施，做為後評估，是無法真正得到最佳的建築環境狀態。藉由人工智慧此項工具，能夠縮短以及應用更少的電腦資源得到評估結果，讓環境評估能夠在設計初期就得以被考量，達到環境性能效益最佳化的目的。而於流程中，最後會以遺傳演算法，搭配訓練完成的深度學習框架去尋求建築生成模型的最佳解。本研究所針對的環境性能模擬對象包括：日照時數、日輻射、熱舒適度、能耗以及總光伏，此外也包含風場模擬。研究中使用深度學習框架去完成預測環境性能模擬指標的任務，判斷城市模型或大樓模型，選擇使用ANN或3D-CNN來預測日照時數、日輻射、熱舒適度、能耗以及總光伏的平均數值。而風場模擬預測任務中，為了測試全域數值的預測效能，論文中使用3D-Unet進行預測。依據不同的預測任務性質與預測對象狀態，選擇不同的深度學習框架才足以正確發揮此工具的優勢。而本研究根據所預測的建築模型是城市模型或者是大樓模型而賦予不同的執行流程。

關鍵字

人工智慧；環境性能模擬；機器學習；最佳化；深度學習；遺傳演算法

並列摘要

Artificial intelligence flourishes thanks to the continuous progress of science and technology. The operation of today’s society has been partially entrusted to artificial intelligence, with the help of which labor costs can be significantly reduced and the efficiency of task execution can be increased. As the application threshold of artificial intelligence tools has been lowered, the work that this tool can take charge in is becoming increasingly common, which makes all walks of life gradually consider how to combine their respective fields and majors with artificial intelligence. With the complex composition of buildings, extensive factors considered and expensive costs, researchers with an architectural background are required to study the combination of architecture and artificial intelligence. In this study, the impact of architecture on the surrounding environment and society should not be underestimated, and there are still a large number of thorny problems to be solved. Therefore, this research attempts to solve the problems related to architecture and the building environment with the aid of artificial intelligence. Building environment is an issue of great concern to architects. It is feasible to evaluate whether the designed building block or equipment system can meet the use standard via environmental performance simulation. However, the current environmental simulation system requires a quite long time and computer resources to calculate environmental value. In these circumstances, the environmental assessment of the building is implemented in the later stage of design. As a post-assessment, it is impossible to truly get the best building environmental state. With the resort to artificial intelligence, it is possible to shorten and apply fewer computer resources to get the evaluation results, so that environmental evaluation can be considered at the early stage of design, further achieving the purpose of optimizing environmental performance and benefits. Finally, the genetic algorithm will be combined with the deep learning framework completed by training to find the best solution for the building generation model. The environmental performance simulation objects for this study include sunlight hours, radiation, thermal comfort, energy consumption, total photovoltaic, as well as wind field simulation. In the research, the deep learning framework is adopted to complete the task of predicting environmental performance simulation indicators and judge the city model or single building model. ANN or 3D-CNN is chosen to predict the average value of sunlight hours, radiation, thermal comfort, energy consumption and total PV. To get the predicted value of the whole base, 3D-Unet is adapted for prediction in the task of wind field simulation. According to the attribute of different prediction tasks and the condition of prediction objects, choosing a suitable deep learning framework is helpful to give full play to the advantages of this tool. This research gives different execution processes according to whether the predicted building model is a city model or a building model.

並列關鍵字

Artificial Intelligence ； Environmental Performance Simulation ； Machine Learning ； Optimization ； Generic Algorithm ； Deep Learning

參考文獻

Google Scholar

英文文獻

Google Scholar

BohnackerH.e.a. “Generative Gestaltung.” Verlag Hermann Schmidt Mainz, 2009.

Google Scholar

ChaillouStanislas. AI + Architecture Towards a New Approach, 2019.

Google Scholar

Danil Nagy, Damon Lau, John Locke, Jim Stoddart, Lorenzo Villaggi, Ray Wang, Dale Zhao and David Benjamin. “Project Discover: An Application of Generative Design for Architectural Space Planning.” Simulation for Architecture and Urban Design, 2017.

Google Scholar

主題瀏覽