Performance Assessments of Clustering-Based Methods for Smart Data-Driven Building Energy Anomaly Diagnosis

The wide adoption of building automation system has collected massive amounts of building operational data, which are of great value to facilitate the decision-making of building professionals. In the past few years, many data-driven approaches have been proposed for building energy anomaly diagnosis. Existing studies mainly utilized clustering analysis as the analytical tool as it can be applied with little prior knowledge. One of the most challenging problems is the performance assessment of clustering-based methods for building energy anomaly diagnosis, as there is no ground truth for validations. This study aims to quantitatively assess the effectiveness of different clustering algorithms in building energy anomaly diagnosis. To ensure the research validity and generalization performance, building energy data from 10 primary schools have been adopted for analysis. Manual labeling has been conducted to provide ground truths on building energy anomalies. A number of data-driven methods have been proposed for identifying daily energy anomalies using different feature extraction and clustering methods. The method effectiveness has been tested using the manually labeled data. This study helps to quantify the value of clustering-based methods in building energy anomaly diagnosis. The research outcomes are beneficial for the development of data-driven methods for smart building energy management.