A Basic Evaluation Criterion of Data-driven Building Energy Model Transferability Assessment
Open Access
- Author:
- Ning, Xiaoyue
- Graduate Program:
- Architectural Engineering (MS)
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- March 20, 2023
- Committee Members:
- Yuqing Hu, Thesis Advisor/Co-Advisor
James Freihaut, Committee Member
Julian Wang, Professor in Charge/Director of Graduate Studies
Greg Pavlak, Committee Member - Keywords:
- Building energy consumption
Transfer learning
Data-driven model
Energy prediction - Abstract:
- Buildings are responsible for 40% of global energy consumption and 33% of greenhouse gas emissions, so it is vital to reduce building energy consumption to achieve sustainability objectives. The key to increasing energy efficiency in buildings is to forecast building energy consumption, which can analyze the energy-saving potential of buildings and guide future energy use, as well as improve the efficiency of energy-consuming building equipment. In recent years, machine learning techniques have been frequently applied to forecasting building energy consumption and performance because of their simple method of operation and high computational efficiency times. However, many buildings do not have adequate data to build accurate energy prediction models, specifically for low-income areas and newly constructed buildings. It can be beneficial if models based on existing buildings that have collected a lot of data can be transferred to buildings with insufficient data by transfer learning. Even though many studies discussed transfer learning applications in different scenarios for building energy prediction, the generalization ability of the transfer learning models hasn't been fully explored. Moreover, there is no well-recognized strategy or principle on how to select a suitable source building. To answer these questions, this study assesses the model transferability across buildings with different building features and in different climate zones (Climate Zone 1A, 4A and 7 are considered in this study) based on the EnergyPlus prototype building model. Specifically, this study selects a medium office building as the baseline model and then adopts the control variable method to change critical building features, including building envelope, HVAC system, and building schedules to prepare the new building's configuration. EnergyPlus is used to run hourly energy simulations for data preparation. Then, this study uses Long Short Term Memory (LSTM) to build a time series machine learning model for baseline building and make a transfer learning model between the original building and the feature-changed building and between baseline buildings in different climate zones. Based on model performance comparison, this study find the impact of building features on model transferability varies cross climate zones. Specifically, the building envelope and HVAC system do not have much significance to the transfer learning performance in Climate Zone 1A, while changing building schedules may cause a decrease in transfer learning performance. In Climate Zone 7, a different HVAC system can result in a negative transfer. For Climate Zone 4A, any change in the building envelope, HVAC system, or building schedule could lead to a negative transfer. For transfer learning between the buildings in different climate zone, there are still some examples that show increased performance. When a building in the climate zone with a more evenly distributed temperature serves as the source building, and the target building is located in a climate zone with great variation throughout the year, a negative transfer occurs. On the contrary, when a building in the climate zone with a more evenly distributed temperature serves as the target building, and a building in the climate zone with great variation throughout the year is the source building, transfer learning shows increased performance.