Machine Learned Façade : Experimental Investigation of AI Model Performance for Daylight Control Using XAI
머신-런드 파사드: XAI를 이용한 일조조절을 위한 AI모델 성능 실험 분석
- 주제(키워드) adaptive architecture , kinetic facade , artificial intelligence , AI , Glare
- 주제(DDC) 720
- 발행기관 아주대학교
- 지도교수 Hwang Yi
- 발행년도 2023
- 학위수여년월 2023. 8
- 학위명 석사
- 학과 및 전공 일반대학원 스마트융합건축학과
- 실제URI http://www.dcollection.net/handler/ajou/000000033003
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
The façade of a building critically influences occupant visual and thermal comfort interfacing between the indoor and the external environment, and dynamic transformation of the façade can improve the building’s performance significantly. This study addresses the use of artificial intelligence (AI) to operate climate-responsive kinetic façade (KF) systems. The KF is a dynamic and morphologically responsive mechanism designed to improve environmental building performance (EBP). Despite recent design advancements and emerging material approaches in the KF building studies, intelligent strategies of morphological operation have been little investigated. To examine our primary hypothesis that optimal undulation with open and close of the adaptive KF controlled by AI models effectively improves indoor daylight glare and illumination performance, we selected three machine-learning (ML) regressors (Extreme Gradient Boosting (XGBoost), Random Forest (RFR), Decision Tree (DTR). The ML models were trained and tested with a set of 20,000 data points from EBP simulation using Radiance on Rhino Grasshopper. KF modules were parametrically designed and fully fabricated (width = 1.73m, height = 1.1m). For demonstration, the KF design and AI control were implemented in a test mockup (floor length = 3m, width = 2m) in Suwon, South Korea. The façade shape was optimized according to indoor daylight probability (DGP) and daylight, respectively in real-time, by using a differential evolution algorithm. Sensor data were collected in dynamic operation. Results indicated that tree-based algorithms outperformed other ML algorithms, achieving more than R2 of 0.85 – 0.90 accuracy compared to neural networks’ R2 values of 0.5 accuracies in predicting DGP and illuminance. Furthermore, we observed that training of a full-scale façade module was highly dependent on the amount of dataset and hyperparameter combinations, and the accuracy of the RFR depended on the maximum depth.
more목차
Chapter 1 - Introduction 1
1.1 Background and Research Questions 1
1.2 Research Objectives 4
Chapter 2 – Literature review 5
2.1 Kinetic Façade and Operation Methods 5
2.2 AI in Architecture and Building Studies 15
2.3. Explainable Artificial Intelligence 18
Chapter 3 – Method and Materials 20
3.1 Façade Design and Experiment Setup 20
3.2 Performance Indicators 27
3.3 Simulation and Data Collection Method 30
3.4 Studied algorithms and model settings 32
3.5 Model Performance Measures 35
Chapter 4 – Results and Discussion 36
4.1 Comparison of ML Models 36
4.2 Analysis of DGP 39
4.2.1 Experimental results of DGP 39
4.2.2 Performance Comparison of ML Models for DGP 44
4.2.3 Performance Comparison Based on Sample Size and Experiment Date for DGP 45
4.3 Analysis of Lux 50
4.3.1 Experimental results of Lux 50
4.3.2 Performance Comparison of ML Models for Lux 56
4.3.3 Performance Comparison Based on Sample Size and Experiment Date for Lux 58
4.4. XAI analysis 62
4.4.1 SHAP Analysis of ML Models for DGP Prediction 62
4.4.2 SHAP Analysis of ML Models for Lux Prediction 69
4.5 Discussion 76
Chapter 5 – Conclusions 82
References 84
