Date of Award

5-10-2026

Date Published

June 2026

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Mechanical and Aerospace Engineering

Advisor(s)

Bing Dong

Subject Categories

Engineering | Mechanical Engineering

Abstract

Buildings account for approximately 40% of total energy consumption in the United States, with HVAC systems representing the single largest energy end-use. Advanced control strategies such as model predictive control (MPC), reinforcement learning (RL), and differentiable predictive control (DPC) have demonstrated significant energy savings in individual studies, yet their widespread adoption remains limited by the lack of scalable, control-oriented dynamic models, the absence of consensus on control strategy selection, and the gap between single-zone research and multi-zone real-building deployment. This dissertation addresses these barriers through an integrated research program spanning physics-informed machine learning (PIML) model development, advanced control evaluation, and scalable multi-zone deployment. For single-zone applications, a physics-consistent input convex neural network (PCICNN) is developed that embeds temporal causality through partial connectivity, ensures input convexity for optimization, and enforces physical consistency through architecture design and loss function augmentation. The PCICNN achieves prediction accuracy within sensor error bounds while maintaining physically consistent behavior under perturbed inputs. Using this model, the first long-term real-building comparison of MPC, RL, and DPC is conducted over five months in the same building, revealing that RL provides the most robust savings (up to 48%) across varying conditions, MPC achieves the highest peak savings (over 50%) but suffers from information gaps and non-optimal decisions, and DPC offers the fastest inference but the lowest overall savings (30.6%). A dual-phase benchmarking framework further establishes PCICNN as the most control-oriented PIML architecture among three representative models evaluated in both prediction and closed-loop control settings. For multi-zone applications, a physics-consistent graph neural network (PCGNN) replaces manual adjacency-matrix neuron connections with group-shared multi-scale causal convolutions and heat diffusion graph layers, enabling scalable multi-zone temperature prediction. A κ-neighborhood truncated critic design for multi-agent RL reduces critic input dimensionality from scaling with the total number of zones to scaling with the local neighborhood size. The framework is validated through a 6-zone simulation and an 18-zone real-building deployment achieving 35–70% energy savings, representing one of the largest real-building multi-agent RL deployments for HVAC control in the literature. As a complementary exploration, a feasibility assessment using the WHO’s comprehensive airborne infection risk model examines whether HVAC-based clean air delivery warrants inclusion as a primary control objective. The analysis demonstrates that when infector presence is known, behavioral interventions outperform ventilation by 3–10 times, making ventilation a supportive rather than a dominant strategy; when unknown, ventilation is the only continuously available strategy and serves as the primary defense. This research contributes to the field through: 1) a control-oriented PIML architecture (PCICNN) for single-zone building dynamics; 2) the first long-term real-building comparison of three advanced control strategies; 3) a dual-phase PIML model benchmarking framework; 4) a scalable PIML architecture (PCGNN) for multi-zone dynamics; 5) a scalable model-based multi-agent RL framework with neighborhood-truncated critics validated through real-building deployment; and 6) an exploratory evidence-based assessment of HVAC’s role in infection risk mitigation.

Access

Open Access

Download

Share

COinS