Graph-Based Visual Analysis for HVAC System Anomaly Detection

HVAC systems are essential components in modern buildings, processing multivariate time series data such as chilled water supply temperature, relative humidity, and indoor thermal comfort parameters to optimize operations. However, detecting abnormal events, including system failures and attacks in such data, is a significant challenge. Existing anomaly detection methods suffer from limitations such as ignoring relationships between variables, failing to identify the causes of anomalies, and lacking interpretability. We propose a correlation and temporal synergy graph-based anomaly detection method that learns the inter-variable and temporal relationships to predict the future values of variables. We first utilize a discrete dynamic graph to model the correlations between variables, and then employ an LSTM-based framework to learn the temporal dependencies of variables in time windows. Anomalies are detected by comparing predictions with actual observations. To help users perceive the occurrence of anomalies and enhance the interpretability of the method, we implement a visual analytics system. The system uses structural differences between subsequent snapshots in the dynamic graph to help users locate anomalous variables and employs a prompt template to provide textual explanations and solutions for the anomalies. The quantitative evaluation shows that our anomaly detection method consistently achieves a precision exceeding 90 %, outperforming baseline approaches. A case study further demonstrates the effectiveness of the visual analytics system in explaining anomalies.