Machine-Learning-Based Model for Supporting Energy Performance Benchmarking for Office Buildings

Abstract

Buildings are dominant contributors of global energy consumption. Enhancing building energy efficiency has long been recognized as an important way to achieve energy saving goals and sustainability targets. In response, a body of building energy performance benchmarking models and tools have been proposed during the past decades. The degree of similarity between the compared buildings is the core of the benchmarking process. However, existing benchmarking tools mainly classify buildings only based on building use types, instead of fully considering a wider range of impacting factors. To address this gap, this paper proposes a machine-learning (ML)-based model for classifying buildings—based on building characteristics, occupant behaviors, and geographical and climate features—into three energy-consumption levels: low, medium, and high. Support vector regression models are then fitted to define the predicted energy consumption for benchmarking. The proposed ML-based building energy consumption prediction model was tested on the office buildings in the commercial building energy consumption survey (CBECS) dataset. Principal component analysis (PCA) was used for data dimensionality reduction and feature extraction. Different ML algorithms were tested and compared, including Naïve Bayes (NB), support vector machines (SVM), decision trees (DT), and random forests (RF). The classification algorithms were evaluated in terms of precision and recall; the regression models were evaluated in terms of root mean square error; and the energy consumption prediction results were further compared with the prediction results by EnergyStar. The performance results indicate that, compared with EnergyStar, the proposed model can reduce the prediction error by 13%.