Open Collections

Wang, Rui

University of British Columbia

Building energy consumption accounts for approximately 40% of global energy use and continues to rise by about 3–4% per year. Consequently, improving building energy efficiency is crucial for a sustainable society. Although recent studies have developed various deep-learning models for building energy load forecasting, many ignore multi-task collaborations, system reliability, and computation efficiency. Therefore, this research aims to develop multi-purpose, distributed data-learning frameworks that enhance the reliability, efficiency, and predictive performance of energy data for community buildings. First, building energy data analytics is extended with more prediction capabilities of energy anomaly detection and anomaly prediction. A new multi-task learning framework is re-invented with encoder-decoder architecture and model optimizations, which significantly improves prediction accuracies of energy load forecasting. Building system maintenance is enhanced with accurate anomaly prediction so early alerts are available. Second, federated energy data analytics are enhanced by improvements in system reliability and prediction accuracies. An adaptive weighting method is designed to mitigate the adverse effects of system faults, and a new deep learning model is designed to improve the forecasting accuracy. Furthermore, an unsupervised anomaly prediction pipeline is developed despite the absence of explicit anomaly labels. Extensive experiments show that this approach achieves high anomaly prediction accuracy, demonstrating the system’s effectiveness. Third, a new personalization strategy is proposed to address the heterogeneous energy data forecasting. Through the sparse selection of new deep learning models, energy load forecasting accuracy and computation efficiency are both improved. Extensive experiments are conducted with different federated training algorithms and this new design outperforms state-of-the-art methods over the energy data forecasting of university campus buildings. The research targets several objectives: 1. Extend the capabilities of building energy data analytics to heterogeneous prediction tasks which support better building energy system maintenance; 2. Enhance the reliability and accuracy of the distributed energy load forecasting with robust and personalized designs; 3. Elevate the energy prediction models to distributed edge applications with higher accuracy and fewer computation resources. With the promising experimental results of real energy data, sustainable and efficient building management is easier to achieve.

Text

2025-02-07

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/90347

Electrical Engineering

University of British Columbia

UBCO

http://creativecommons.org/licenses/by-nc-nd/4.0/