Bootstrap-LOCI data mining methodology for anomaly detection in buildings energy efficiency

Abstract

An automated methodology is proposed to identify anomalies in buildings’ HVAC systems, through Local Correlation Integral (LOCI) algorithm, improved by Bootstrap to obtain a rule from its score distribution. It has been performed to solve the case study of anomaly detection for HVAC facilities maintenance in a clothing store in Panama. It is defined by a dataset composed of 24 daily readings recorded during 434 days. In each reading, 15 critical quality variables for thermal comfort and energy efficiency were monitored. For algorithm training, anomalous events recorded by HVAC system operators are considered. In this stage, the LOCI parameters that best fit the data are estimated, to obtain a score for each of the observations and then study their distribution by applying Bootstrap techniques to improve the classification performance. For the algorithm performance evaluation, cross-validation is used and from these results, it is compared with two benchmark supervised classification methods such as Logistic Regression and Support Vector Machines with polynomial kernel. The LOCI method has been improved by the application of bootstrap resampling, providing estimates of the LOCI score distribution and a critical value from which we define an anomalous observation. It provides the best balance between real anomaly detection and prevention of false alarm identification than current proposals for LOCI.

Keywords

• LOCI
• Anomaly detection
• Bootstrap
• Energy efficiency