An interpretable method for identifying mislabeled commercial building based on temporal feature extraction and ensemble classifier

Abstract

• The model is developed based on temporal feature extraction and attempted to maintain temporal features. • The model is developed based on supervised learning and aims to identify mislabeled building based on the current building categorization method. • The model is developed with a good interpretation to face practical engineering problems. • Discuss the mislabeled buildings in reality and the temporal differences among different PSU-type buildings. • Discuss the renewal or improvement of PSU categorization. Proper building categorization is important in building energy efficiency analysis. Primary space usage (PSU) is a typical and widely used commercial building categorization method. The PSU labels are ascertained once the buildings are put into use but not always modified on time when the building usages change, which may lead to false results in analysis. In this paper, we propose a method to identify mislabeled commercial buildings based on analysis of the energy time series collected by electric meters. The method is constructed as follows: (1) data cleaning and transformation; (2) three types of temporal feature extraction; (3) several single classifier training, and the ensemble classifier building; (4) mislabel building identification and correction. The method provides a supervise way to identify mislabeled building. We applied the method to a public dataset from the Department of General Services from Washington, D.C. and found that 22.4% of the buildings were mislabeled. We also designed 1000 evaluation cases to prove the effectiveness of the method. Based on the results of the cases and the good interpretation of the method, we discuss the mislabeled buildings in reality and the temporal differences among different PSU-type buildings. We also discuss the renewal or improvement of PSU categorization.