Multi-model real-time energy consumption anomaly detection for office buildings based on circuit classification

With the widespread adoption of office building electricity consumption monitoring platforms, ample data are available for diagnosing energy anomalies, increasing interest in data-driven approaches. However, wholebuilding energy evaluation often fails to identify anomalies in specific sub-circuits. Additionally, the complexity of building energy systems has led research to focus mainly on data-driven methods, with limited exploration of individual sub-circuit characteristics. To address these issues, this study proposes a classification procedure based on physical attributes and data features of office building power circuits, categorizing energyconsumption circuits into four types. Subsequently, a multi-model real-time diagnostic framework was developed, which utilizes anomaly detection models tailored to specific circuits for precise identification of anomalies. The framework was experimentally validated using real-world data from a commercial office building in Haidian District, Beijing. The results demonstrated that the proposed method effectively performed hourly monitoring of energy consumption in lighting, chiller, and cooling tower circuits, and successfully identified multiple time periods during which energy consumption deviated from the normal range due to improper operations by facility management personnel. These findings highlight the benefit of integrating sub-metering with data mining, providing building operators with a novel approach to swiftly detect circuit-level abnormalities and optimize energy management strategies.