Human-centered Ambient Artificial Intelligence In Smart Buildings

Abstract

This dissertation advances human-centred Ambient Artificial Intelligence (AmI) by investigating three dimensions of context awareness, including user persona, building operations, and human states, as key components of an envisioned integrated smart built environment framework. First, a nationwide survey (N =307) quantified how demographics, prior smart home experience, and personal values shape users adoption of smart devices and their acceptance of smart building automation levels. A recommender system employing machine learning models was studied and achieved 92% accuracy and F1 score in predicting users' preferred configurations and control modes, demonstrating that user persona-driven modeling can inform adaptive configuration of smart buildings. Second, a large language model-enhanced Building Energy Management (BEM) AI agent was studied as an integration of perception, reasoning, and action modules. Evaluated on a proposed benchmark comprising 120 natural-language queries across four residential datasets, the agent achieved 86% accuracy in device control, 74% in device scheduling, 77% in multi-step energy analysis, while indicating potential for improvement in cost management (49%). Furthermore, statistical analyses showed that the proposed AI agent framework generalizes well across different buildings. These results establish a benchmark for LLM-based energy agents and highlight trade-offs between reasoning depth and computational resource requirements. Third, a comprehensive dataset—comprising two laboratory-based datasets and one real-world field dataset with varied data collection complexities—was used to investigate wearable sensor–based detection of stress as a complex human state. Eight classical machine learning classifiers and four state-of-the-art LLMs were compared using pre-processed six different biosignal inputs. On the WESAD laboratory dataset, machine learning models achieved up to 96%/94% (Accuracy/F1), while GPT-4.1 reached a comparable 91%/87% using few-shot prompting. For the more complex WDDISSE lab dataset, Gradient Boosting attained 82%/78%, and GPT-4.1 followed with 79%/73%. In the field-based WorkStress3D dataset, SVM performed best at 78%/76%, with Gemini-2.5-Pro at 74%/74%. In cross-dataset lab-to-field generalizability analyses, a fine-tuned Gemini-2.5-Flash model achieved 78%/77% performance, outperforming other cross-dataset ML and LLM baselines. The findings suggest that using statistical summary features and z-score scaling data pre-processing methods, as well as prompt engineering (Chain-of-Thought, Few-shot) and fine-tuning, can consistently improve LLM performance. Taken together, the findings on these foundational components inform future research directions for a context-aware, human-centric, and embedded Ambient Intelligence (AmI) system. The data compilation and processing methods, proposed modular architectures, and evaluation protocols studied in this dissertation may provide a reproducible foundation for future research on context-aware Ambient Artificial Intelligence in smart built environments.