A Variational Graph Autoencoder Aided Canonical Correlation Analysis-Based Online Abnormal Patterns Detection Method for Buildings HVAC Systems

Heating, Ventilation, and Air Conditioning (HVAC) systems have become essential components of contemporary life, extensively employed in commercial and residential settings to ensure environmental comfort. However, worn components and suboptimal settings affect their operation, causing them to operate in abnormal conditions and leading to excessive energy consumption. The operating status of HVAC also fluctuates as load demand fluctuates, further complicating accurate anomaly detection. The advent of the Internet of Things technology has elevated the intelligence of HVAC systems, concurrently providing a robust data foundation for implementing data-driven methods in anomaly detection. Nonetheless, traditional data-driven detection methods often produce high false alarm rates when the system’s operating modes change. This paper proposes a variational graph autoencoder aided canonical correlation analysis (CCA) based online abnormal patterns detection method. The main feature of this method is to analyze the correlation of historical data, thereby building a reconstruction graph that identifies and selects the most relevant neighboring nodes in the graph for the local modeling of current query data. On this basis, CCA is used for online anomaly detection. This method solves the problems of unclear cluster boundaries, difficulty determining the number of cluster centers, and more false alarms when working conditions change in traditional methods. Experimental comparative analysis using HVAC system data from actual buildings proves this method is superior to existing solutions regarding anomaly detection accuracy, false alarm rate, and false negative rate.