運用AI機器學習模型預測廳堂初始空間設計之殘響時間研究

Compare Reverberation Time Prediction with Machine Learning Model and Traditional Physical Formulas in the Initial Stage of Performance Hall Design

初期設計之決策，對整體建築效能影響極大，於廳堂設計亦是。為了嘗試解決傳統建築物性能模擬評估效率不佳的弊病，本研究通過開發有效之AI機器學習模型來預測廳堂初始空間設計之殘響時間，以輔助建築師與聲學顧問，做初期設計決策參考。本研究運用文獻所知31個重要既有廳堂邊界設計之幾何參數，先採用主成分分析來識別影響空間聲學之主要影響幾何特徵，並接續使用機器學習等工具，如分別運用線性迴歸（LR）、分類和回歸樹（CART）、人工神經網絡（ANN）和支持向量回歸（SVR）四種單一模型，以及對應之複合集成模型，即投票Voting、套袋Bagging和堆疊Stacking三種模型來預測廳堂殘響時間。
分析結果顯示，廳堂長寬比例、座椅淨面積、舞台至最遠座位之距離、平均側牆寬度、平均天花板高度、側牆角度對殘響時間影響較大，輸入模型訓練後，用以預測大量的參數化設計模型，以採用LR作為弱學習器之Bagging複合模型，對照於Sabine公式計算殘響值，得到最佳的預測，兩者之相關係數（R）為0.95，平均絕對百分比誤差（MAPE）為18％，均方根誤差（RMSE）為0.52（sec），平均絕對誤差（MAE）為0.44（sec）。期望在資訊極度匱乏的初始設計狀態，能做到可接受誤差內之預測。

The decisions made during the initial design phase are essential to the building's performance, as is the design of the performance space. In order to solve traditional building performance simulation drawbacks, such as the difficulty to get sufficient information and time-consuming detail 3d model construction, an effective workflow combining AI machine learning algorisms to predict design performance is proposed in this study. The predicted values generated by the AI training model and the predictions of traditional physical models will both compare with the actual data of the onsite measurement to get the error rate comparison. We take the reverberation time, a well-known architecture acoustic index, to predict the initial acoustic situation of performance hall design. Sabine, Eyring, and Arau-Puchades formulas embedded in Odeon were used as the traditional physical formula to predict the reverberation time. Linear regression (LR), classification and regression tree (CART), artificial neural network (ANN), and support vector regression (SVR) are the single AI machine learning models used here. Combined with the single model, this research also used three ensemble models, using voting, bagging, and stacking methods to better predict results. Compared with the traditional physical formulas predictionand the machine learning prediction result, we obtained the best machine learning prediction with a mean absolute percentage error (MAPE) of 12%, a root mean square error (RMSE) of 0.395 (sec). The mean absolute error (MAE) is 0.27 (sec). It is expected that the prediction can achieve within the acceptable error in the initial design stage of performance space design, where information is extremely scarce. This method can significantly help the cooperation between architects and acoustical consultants. The proposed workflow can view as a new prediction method for further architecture acoustical studies to understand the limitation and possibilities of AI models comparing with the traditional simulation methods.