Dissertation Defence: Buildings Energy Data Analytics with Multi-task and Federated Learning
January 21, 2025 at 8:00 am - 1:00 pm
Rui Wang, supervised by Zheng Liu, will defend their dissertation titled “Buildings Energy Data Analytics with Multi-task and Federated Learning” in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Electrical Engineering.
An abstract for Rui Wang’s dissertation is included below.
Examinations are open to all members of the campus community as well as the general public. Please email zheng.liu@ubc.ca to receive the Zoom link for this exam.
ABSTRACT
Building energy consumption accounts for 40% of global energy usage and is growing at 3-4% per year. Improving energy efficiency in buildings is crucial for sustainable development. This research develops multi-purpose distributed data learning frameworks to enhance the reliability, efficiency, and prediction performance of energy data in community buildings. Firstly, a multi-task framework addresses four tasks: energy load forecasting, anomaly detection, anomaly prediction, and air temperature forecasting. Using a mixture-of-experts architecture, the feature extraction and prediction network are decoupled to support various prediction tasks. Self-attention mechanisms and linear approximation improve deep learning computation efficiency, with experiments showing significant accuracy improvements. Secondly, an adaptive federated learning system addresses data isolation and system interruptions. By compensating for missing models and using RNN with dilations and CNN for parallel processing, multi-step energy predictions improve efficiency. An unsupervised clustering method enhances anomaly prediction without predefined labels, demonstrating significant forecasting and anomaly prediction improvements. Thirdly, a novel approach applies probabilistic ensemble architecture for personalized energy load forecasting, using a transformer model for increased efficiency. Experiments show the effectiveness of personalized federated learning algorithms in achieving state-of-the-art prediction accuracies for university campus buildings. The research aims to enhance federated learning reliability, improve prediction accuracy and efficiency, extend model capabilities, and scale federated training across different energy data distributions. Promising experimental results suggest more sustainable and efficient building management, reducing energy and maintenance costs.
