| Author: | Xie, Lingyun |
| Title: | AI-enabled smart field control optimization for building central cooling systems |
| Advisors: | Wang, Shengwei (BEEE) |
| Degree: | Ph.D. |
| Year: | 2026 |
| Department: | Department of Building Environment and Energy Engineering |
| Pages: | 1 volume (various pagings) : color illustrations |
| Language: | English |
| Abstract: | Rapid advances in Artificial Intelligence (AI) are reshaping energy-intensive industries, offering transformative opportunities to optimize complex engineering systems for both sustainability and performance. Given that building HVAC systems account for a large portion of global energy usage, researchers have increasingly explored AI-enabled strategies to enhance intelligent facility management and advance energy conservation goals. However, practical deployment in real-world environments remains constrained by key barriers, including high computational demands that limit real-time feasibility, reliance on centralized architectures vulnerable to communication bottlenecks, insufficient robustness under system uncertainties, limited adaptability to equipment degradation, and scalability challenges in complex multi-zone applications. This PhD research addresses the limitations of current HVAC optimization practices by proposing a generic AI-empowered framework and a suite of innovative control strategies—including lightweight online optimization, energy-gradient-guided decision-making, incremental online learning, and parallel distributed architectures—thereby enabling efficient, robust, and scalable real-time deployment of intelligent optimization in building central cooling systems. A comprehensive review is conducted to examine the progression of HVAC control strategies, ranging from traditional rule-based methods to contemporary AI-enabled approaches. The review reveals notable progress in the theoretical development of intelligent control strategies. It also reveals several persistent challenges that limit the real-world applicability of current strategies. A notable lack of autonomous, interpretable, and resilient AI mechanisms underscores the gap between theoretical advances and practical implementation. These insights define the research objectives and guide the design of the proposed strategies. To enable AI-enabled field-level optimization, a generic framework and associated strategies are developed, integrating AI-based control optimization into BAS by adopting smart control stations at the BAS field level. The proposed framework comprises two core functional modules: (1) an AI operating environment that supports lightweight and real-time execution of AI models, and (2) a comprehensive suite of AI functional boxes that ensure effective and reliable execution of AI algorithms. The framework is validated by integrating two physical smart control stations with a BAS testbed. Control robustness and energy performance are further assessed through hardware-in-the-loop testing with a simulated dynamic HVAC system. Test results show that the proposed framework and associated strategies can maintain stable and robust control under various critical conditions. Building on this foundation, the subsequent optimization strategies are also validated on the same hardware-in-the-loop testbed using smart control stations. To address the challenge of computational burden in real-time optimization, a simplified deep learning-enabled Genetic Algorithm is proposed, ensuring optimization intervals are short enough for online applications. This algorithm also significantly reduces CPU and memory usage, enabling deployment on a miniaturized control station for field implementation. To further enhance stability and reliability, a robust assurance scheme is incorporated, which switches to expert knowledge-based control under abnormal conditions. Hardware-in-the-loop testing validates the proposed strategy, confirming its computational efficiency, control effectiveness, and operational robustness. Test results show that the optimal control strategy achieves an average of 7.66% energy savings and exhibits strong operational robustness. To enhance robustness against data uncertainties, a novel local energy gradient-guided AI-empowered online optimization strategy is proposed. Relative energy consumption in neighboring regions is employed to guide the optimization process, thereby improving control stability and enhancing resilience to input data noise. This strategy involves theoretically formulating a physical variable, the "energy gradient", to guide the AI-empowered online optimization method. A machine learning model is employed to capture the trend of the energy gradient, guiding the progressive adjustment of setpoints. A comprehensive system mechanism analysis is conducted to uncover system dynamics and provide theoretical support for the proposed strategy. Results show that the proposed strategy exhibits superior stability and robustness, with an energy-saving effect comparable to those achieved by conventional online optimization strategies, achieving approximately 5% energy savings. To improve long-term adaptability under system degradation and dynamics, an incremental online learning strategy using structural augmentation of the optimization model is introduced. A residual neural network branch is added alongside the offline-trained hybrid model and selectively updated online, enhancing adaptability while preserving pre-trained knowledge. A low learning-rate Adam optimizer is adopted to continuously update the model using real-time data, maintaining stability and preventing catastrophic forgetting. Test results show that the incremental online learning approach effectively restores system performance after degradation, reducing energy consumption by 5.77%, while the original optimization algorithm achieves only 1.12% savings under degraded conditions. To enhance system-level scalability and fault tolerance, a parallel distributed optimal control strategy is developed. Based on a decentralized operational framework, agents spatially distributed across field stations or edge devices can perform optimization tasks independently. By leveraging dynamic task partitioning, peer-to-peer communication, and voting-based consensus mechanisms, these agents collaboratively achieve global system optimization without the need for centralized coordination. Hardware-in-the-loop tests are conducted by deploying the strategy on four real smart control stations. Test results show that the proposed strategy improves scalability, operational reliability, and optimization efficiency, while delivering 9.16% energy savings. In conclusion, the framework and strategies developed in this PhD research bridge the gap between academic innovation and practical deployment of AI-enabled control optimization in large-scale building cooling systems, offering a scalable, reliable, and energy-efficient pathway toward intelligent building operations. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/14428