Data-driven prediction and optimization in building energy consumption

• Main Author: Tan, Yong Shen
• Other Authors: Zhang Limao
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Civil engineering
• Online Access: https://hdl.handle.net/10356/154453
• Institution: Nanyang Technological University

The data records for building energy consumption of Seattle City in 2015 and 2016 are obtained from the annual data reports and used as the datasets in data-driven prediction and optimization models’ development for energy efficiency evaluation and enhancement. A gradient boosting regression approach named Light Gradient Boosting Machine (LightGBM) is proposed to achieve accurate prediction of building energy consumption through regression model constructed by learning and training large dimensions of multi-source heterogenous data from the annual reports. In case study of this project, the proposed methodology is used to analyze the Seattle’s building energy performance as well as predict the weather normalized site energy use intensity and the greenhouse gases emissions intensity of properties in the city. The non-linear relationships between each objective function and the influential features are characterized by the algorithm. The prediction results obtained from the LightGBM models are compared with the observed values and the models showed coefficients of determination, R2 of 0.8046 and 0.9029 for Weather Normalized Site Energy Use Intensity and Greenhouse Gases Emissions Intensity respectively. The feature importance of each input is also analyzed to select highly influential features on the objective functions for multi-objective optimization. A multi-objective optimization algorithm approach named Non-dominant Sorting Genetic Algorithm II (NSGA-II) is proposed as the methodology to perform the optimization of both objective functions which are weather normalized site energy use intensity and greenhouse gases emissions intensity of the buildings. A total of twelve influential features and two objectives are identified after the data preprocessing and the NSGA-II algorithm is utilized to develop optimization models using the test dataset, where only the input features with top three highest influencing power are involved in the variable adjustment. After the multi-objective optimization is conducted, the optimal solutions obtained are further analyzed to access the effectiveness of the proposed methodology and an overall improvement percentage of 77.52% is achieved for the objective functions by modifying the three highly influential factors. The equivalent optimal inputs are also investigated as a reference for the building users to manage the energy performance and achieve ideal building energy consumption.