A Robust Unsupervised Framework for High-Resolution Building Energy Consumption Profiling

Unsupervised learning methods have been widely used for building energy consumption profiling, but the currently used methods usually gave undesirable results and could hardly tolerate the highly diversified building cases. A scalable automatic framework is accordingly proposed in this study to achieve accurate load profiling. The framework consists of four major steps: pre-processing, preliminary K-means clustering, DBSCAN within clusters, and post-processing. With a dataset including 50 different buildings in Singapore, the framework was demonstrated to outperform all the baseline methods in most cases (37 out of 50). The profiling results provides more comprehensive insights on the buildings energy behavior, facilitates applications such as building load prediction and improves the prediction accuracy.