|Development of a data-driven methodology for dynamic energy modeling of large buildings towards mass deployment of AI in smart buildings
|Xiao, Fu Linda (BEEE)
|Buildings -- Energy conservation
Buildings -- Energy consumption
Hong Kong Polytechnic University -- Dissertations
|Department of Building Environment and Energy Engineering
|1 volume (unpaged) : color illustrations
|Building operation accounted for 30% of global final energy consumption and 27% of total energy sector emissions in 2021, reported by International Energy Agency (IEA 2022). Large buildings are major electricity consumers of power grid, affecting the demand balance and power grid reliability significantly. Building energy management in the operation stage is crucial for not only energy conservation, but also emission reduction, demand response and improvement of occupancy comfort and productivity. Building energy modeling has proven to be crucial for evaluation and optimization of building operation, thus enhancing the management of building energy systems. However, the energy modeling of large buildings with complex energy systems remains a great challenge for conventional physics-based approaches, which is labor-intensive, time-consuming, and requires detailed information of buildings and building energy systems. As the development of Internet of Things (IoT) and Building Automation Systems (BAS) which provided reliable sources of massive building operation data, the AI technology has attracted increasing attention in various aspects of building energy management. In comparison to traditional physics-based methods, AI exhibits several promising advantages, including increased capability to address nonlinear problems, less requirement on knowledge/information of buildings and building energy systems, ease of implementation, etc. However, there are still some major impediments preventing AI from wide and flexible deployment in smart buildings, including model generalization ability, model interpretability, and model updating, and a significant portion of research is still in the exploratory stage with limited applications in real buildings.
This research aims to address the main barriers in the AI deployment in smart buildings for online building energy management, including data scarcity, model generalizability, interpretability and update issues. Based on a comprehensive review of advanced machine learning techniques and in-depth understanding of building operations, this research develops generalizable and interpretable data-driven methods for modeling the dynamics of large buildings and complex energy systems for improving building operational performance.
Accurate building energy prediction is vital for various building energy management tasks. This research proposes an attention-based sequence-to-sequence day-ahead building cooling load prediction model architecture with enhanced interpretability. The model developed based on this architecture is assessed using real building operational data, and shows improved accuracy and interpretability compared with recurrent neural networks without attention mechanisms and other baseline models using a recursive approach. The visualization of obtained attention vectors shows why predictions are made by the model, as well as the proportional influence of inputs on the output. Further analysis of attention-based recurrent models trained with operational data from different buildings provides crucial temporal information for understanding the building dynamics, like the thermal response of the building. For information-poor buildings with insufficient available data, a transfer learning method is proposed to pre-train the building energy prediction model using data from similar buildings. The effects of transfer learning-based model pre-training on building energy prediction model performance are investigated under different conditions, including the volume of source dataset and target dataset, and the similarity between source building and target building. The research outcomes can provide valuable guidance for the implementation of transfer learning, especially in selecting appropriate source buildings and datasets for developing accurate building energy prediction models.
A design information-assisted graph neural network-based methodology is proposed for dynamic modeling of central air conditioning systems. To capture the structural relationships in the physical configuration in buildings energy systems, this methodology symbolizes the physical entities and their relationships as nodes and edges in a graph based on graph theory, and then uses the graphs as the input of machine learning models. The proposed methodology is tested on historical operation data of the central air conditioning system in a high-rise commercial building in Hong Kong. The dynamic modeling of this air conditioning system is regarded as a node-level time series prediction task. Image identification methods are employed to extract the topology of the air conditioning system from 2D schematic drawings, which are used as the structure of the graph. GCN-RNN and GAT-RNN show improved prediction performance compared with conventional deep learning models. And SHapley Additive exPlanations (SHAP) method is adopted to explain the model output at global and module level. The proposed methodology provides a valuable reference for incorporating physical knowledge of building energy system into machine learning modeling, and demonstrates an ideal synergy between machine learning and domain expertise.
Once the models are deployed in buildings, the data-driven models need to be continually updated to adapt to the changes of building operations to avoid performance degradation over time. The main challenges in model update process include concept drift, catastrophic forgetting, model long-term accuracy and computation resources. This study proposes a continual learning-based model update method, which enables the artificial intelligence systems to adapt to new tasks over time without forgetting its previous knowledge. And a comprehensive investigation is conducted on the performance of three conventional update methods and five emerging continual learning methods, using 2-year data of 100 buildings. The results show that continual learning methods exhibit more powerful capability in ensuring long-term accuracy, e.g., the CV-RMSE of Elastic weight consolidation and Gradient episodic memory decreased by around 14% and 8% on average compared with static model and accumulative learning. And their computation times and data storage costs are reduced as well. The comparison results are valuable to the development of adaptive data-driven building energy prediction models which are more reliable over time and robust against changing operation conditions, thus more practically applicable in smart building energy management.
The proposed data-driven methods are tested for online real-time control of a campus building. The real-time data transmission between Building Management System and the AI engine is realized via BACnet and an Influxdb database. An online data preprocessing module is developed for ensuring the data quality of continuous data stream. A daily peak demand limiting strategy is deployed based on 5-days ahead building cooling load prediction model. The demand limiting strategy is tested on August 2022 and February 2023, and the cooling load prediction model is updated using the proposed continual learning method. The results show that the demand limiting strategy can achieve 542 kVA peak load reduction and 62,330 HKD cost saving in August 2022.
In summary, based on a comprehensive review of advanced machine learning techniques and in-depth understanding of building operations, this research develops generalizable and interpretable data-driven methods to model the dynamics of large buildings and complex energy systems for improving building operational performance. The proposed methods target to address the main barriers in the AI deployment in smart buildings for online building energy management, including data scarcity, model generalizability, interpretability and update issues. And their effectiveness is validated using real building operation data and open-source building dataset.
|All rights reserved
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: