Detecting anomalies in building energy consumption can reduce unnecessary energy waste and improve energy efficiency. The role of anomaly detection has become particularly pivotal in industrial buildings because of their high energy consumption and the potential risks associated with abnormal events. Although extensive data collected through smart meters has indicated the advantages of anomaly detection using data mining techniques

labeled data are often unavailable in practical situations. Therefore

this study develops an ensemble framework that combines three unsupervised learning algorithms

including Local Outlier Factor

Deep Isolation Forest

and Anomaly Transformer

to identify anomalous power consumption with a focus on subsequence anomaly. The transformer-based network is established to precisely impute missing values and enhance the reliability of anomaly detection. The experimental results based on hourly cooling energy consumption in the two industrial buildings confirmed the effectiveness of the proposed method. To better interpret the anomaly detection results

the Extreme Gradient Boosting is applied to construct the relationship between influencing factors and anomalous consumption. The area under the Receiver Operating Characteristic curve is used as a metric for the classification task

and an average of over 0.96 indicates robust performance. Weekday and dew point temperatures are found to have significant impacts on the electricity usage pattern. The research findings provide valuable insights for developing effective solutions to identify unexpected trends in building energy consumption and support efficient energy management.