| Author: | Chen, Zhe |
| Title: | Development of data-driven methods with enhanced interpretability and reliability for optimal control and diagnosis of air-conditioning systems |
| Advisors: | Fu, Xiao (BEEE) |
| Degree: | Ph.D. |
| Year: | 2025 |
| Department: | Department of Building Environment and Energy Engineering |
| Pages: | xiv, 152 pages : color illustrations |
| Language: | English |
| Abstract: | Enhancing the energy efficiency of building energy systems has become a key issue regarding carbon neutrality because buildings consume about 36% of energy and contribute 40% of carbon emissions. In the past decade, machine learning and data-driven modeling have been widely studied for building energy management applications, such as fault detection and diagnosis, optimal control, and demand response. Despite extensive research, the data-driven approach faces two major challenges regarding practical applications: a lack of interpretability and reliability. Interpretability refers to providing more insightful information for respective data-driven applications. Reliability aims to improve the applicability of data-driven model-related applications, such as model generalization ability, robustness and reliability in HVAC controls. Therefore, this PhD study aims to develop data-driven methods with enhanced interpretability and reliability for optimal control and diagnosis in HVAC systems. In this study, the following research questions are addressed using novel methods with enhanced reliability and interpretability: Q1: How can we develop data-driven models with high interpretability and generalization ability under data scarcity? Q2: How can we address the measurement uncertainty issue when developing data-driven models? Q3: How can we resolve the physical inconsistency issue to enhance the interpretability and reliability of data-driven models? Q4: How can we achieve reliable online control (balancing energy efficiency and control smoothness) when deploying data-driven models in complex optimization problems? Questions Q1 to Q3 address general issues in the development of data-driven models, while Q4 addresses the application of data-driven models in complex optimization problems. In CHAPTER 3, the data scarcity issue (Q1) is addressed by proposing a similarity learning-based method for enhanced interpretability and generalization ability, especially in typical HVAC FDD applications where only limited labeled data are available. The output of the proposed model is the similarity of the input to each fault, which provides better interpretability than conventional models that only output the probability of each fault. A temporal data-splitting method is utilized to handle the high correlation between training data and test data when a random split is adopted. The proposed method is validated in two scenarios: insufficient labeled data (scenario 1) and imbalanced labeled data (scenario 2), respectively. In scenario 1, the proposed method improves diagnostic accuracy by 45.7% compared to the baseline model when labeled data are limited. In scenario 2, the method demonstrates superior generalization ability when dealing with imbalanced labeled data, improving fault diagnosis accuracy from 88.4% to 90.8%. In CHAPTER 4, the measurement uncertainty issue (Q2) is addressed by probabilistic machine learning with a risk evaluation scheme for enhanced interpretability and reliability. For the typical chiller sequencing control problem, this study proposes a robust chiller sequencing control strategy to overcome the limitations faced by conventional cooling load predictions in controlling chiller plants. The strategy utilizes probabilistic cooling load predictions, with one-hour-ahead probabilistic cooling load predictions formulated as a normal distribution using the natural gradient boosting algorithm. A risk-based action evaluation scheme is designed to determine the optimal number of operating chillers by assessing the risks of possible control actions. The proposed strategy is validated through case studies based on historical operational data and an in-situ test. The case studies show that the strategy significantly enhances the reliability of the chiller plant by reducing the total switching number of chillers by up to 43.6% (from 55 to 31 in a week). The in-situ test results reveal a 56.5% decrease in total switching number compared to the rule-based strategy. The average daily energy savings is approximately 3945.1 kWh. Additionally, the coefficient of performance of the chiller plant increased by 4.2%. In CHAPTER 5, the measurement uncertainty issue (Q3) is addressed by a new physically consistent data-driven method for enhanced interpretability and reliability. For a typical variable speed pump modeling problem, this study proposes a physically consistent, data-driven optimal sequencing control strategy to address the often-overlooked issue of optimal pump sequencing in HVAC water systems that adopt variable speed parallel pumping. By integrating physical knowledge with data-driven modeling, this approach enhances interpretability and reliability in real applications. The interpretable ElasticNet model is trained using features including pump frequency and total flow rate for different numbers of operating variable speed pumps. The optimal sequencing is subsequently determined using a physically consistent speed/frequency prediction method. The proposed strategy is validated using real building operational data, with results showing an average of 10% energy saving compared to the conventional rule-based strategy. In CHAPTER 6, reliable online control (Q4) is achieved by a novel multi-objective optimization-based framework that enhances smoothness--an aspect often neglected in conventional deterministic and metaheuristic optimization algorithms. The distance between two successive control actions based on Manhattan and Euclidean distances is one optimization objective, along with energy conservation as another optimization objective. The proposed strategy is validated using a typical optimization problem: optimal chiller loading. Results show that the proposed approach significantly improves control stability while maintaining energy efficiency compared to conventional optimization algorithms. To sum up, the proposed data-driven methods improve interpretability and reliability compared to existing data-driven methods, addressing general issues in the development of data-driven models and the application of data-driven models in complex optimization problems related to fault diagnosis and optimal control in central air-conditioning systems. The new methods developed in CHAPTER 4 and CHAPTER 5 (i.e., chiller sequencing strategy and pump sequencing strategy) have been deployed to the real chilled water system in the PolyU campus, demonstrating the energy efficiency and applicability of the proposed methods. In the future, the proposed data-driven methods are promising for large-scale deployment. |
| 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/200/13607