Degree Name

Doctor of Philosophy


School of Mechanical, Materials, Mechatronic and Biomedical Engineering


Buildings account for a significant portion of energy consumption worldwide and are responsible for a considerable amount of carbon emissions. In the pursuit of global energy consumption reduction and carbon neutralization, effective strategies for building energy management play a pivotal role. By leveraging building operational data and data science techniques, data-driven strategies have emerged as a promising method to optimize building energy performance without the need to develop complex building models. This thesis aims to develop novel data-driven strategies to optimize building energy management with a focus on key areas, including building performance evaluation, energy usage prediction, and demand flexibility optimization.

An efficient data mining strategy is first presented to assess the performance of the heating, ventilating and air conditioning system of a residential building cluster. This strategy uses Symbolic Aggregate approXimation and Kernel Density Estimation to identify typical daily load patterns and evaluate overall system performance. A customized association rule mining model was developed to discover the associations among different attributes and identify the root causes of inefficient operations. This strategy was evaluated using one-year operational data of a gound source heat pump system, effectively identifying efficient and inefficient operation patterns and revealing excessive power consumption of water pumps as the main factor impacting energy efficiency. This strategy can also be adapted for system-level performance evaluation of other energy systems with appropriate modifications.

This thesis then presents a predictive modeling strategy to forecast the next-day total and peak electricity demand of a building portfolio, with the main objective to enable accurate building-level performance assessment and facilitate building demand side management. The strategy utilizes long short-term memory (LSTM) models for energy usage prediction and reinforcement learning (RL) agents to dynamically tune the parameters of LSTM models based on their prediction errors. The strategy was tested using electricity consumption data from a group of university buildings and student accommodations. The results demonstrated improved prediction accuracy for buildings with large monthly variations in electricity usage in comparison to that using LSTM models alone.

The thesis further explores the development of novel data-driven strategies to optimize building operations. A framework that integrates machine learning algorithms and a domain knowledge-based expert system is proposed to improve building energy flexibility by using solar photovoltaic (PV) and battery storage systems. The framework utilizes a rule-based expert system (RBES) to maximize PV self-consumption, an RL agent to optimize grid power import for battery charging and discharging decisions, and a Classification and Regression Tree (CART) model to analyze the relationship between building energy flexibility and external variables. The performance of the framework was tested using the four-year data collected from a low-energy office building. The results showed that the integration of RL, RBES and the CART model can result in reduced electricity costs (7.0%), decreased grid power consumption (10.6%), and increased PV self-consumption (9.2%) as compared with the RBSE strategy.

To further improve building energy flexibility across multiple indicators, a strategy integrating an RBES and multiple RL agents is proposed. Unlike the previous strategy, this approach uses the RBES to provide references for RL to explore optimal solutions within a reduced optimization space, improving exploration performance and avoiding unreliable decisions. The proposed strategy was tested using PV generation data and energy consumption data of a net zero energy building. The results showed that the strategy can improve the RL learning efficiency by up to 85.7% and successfully avoid sub-optimal convergence during policy learning. Compared with the rule-based expert system alone, the proposed strategy can reduce the operational cost by 5.4% and the daily peak-to-average ratio of grid power during peak hours by up to 19.2% while maintaining the same level of PV self-consumption. Compared with a model predictive control method, the strategy achieved similar decision performance in cost savings while using significantly reduced decision time.

The data-driven strategies developed in this thesis hold the potential to assess the effectiveness of different building energy systems and support demand-side management within buildings.

FoR codes (2008)


This thesis is unavailable until Thursday, March 05, 2026



Unless otherwise indicated, the views expressed in this thesis are those of the author and do not necessarily represent the views of the University of Wollongong.