A Practical Framework for Energy Fault Detection in Smart Buildings

With the increasing global energy demand, energy-efficient building operations have become more imperative for sustainability. Building energy consumption patterns serve as critical indicators of operational efficiency and system health. Fault Detection and Diagnosis (FDD) is used to automate the process of detecting faults and tracking system health in large-scale buildings. Implementing an FDD system minimizes manual inspections while increasing the reliability of fault detection. However, the absence of historic fault labels makes fault detection a challenging problem. This paper proposes a novel multistage framework that can serve as a blueprint to detect energy faults in smart buildings. The proposed framework comprises three main components: outlier detection and imputation, energy forecast, and fault detection. Using the Large-Scale Energy Anomaly Detection (LEAD) dataset, we evaluated multiple fault detection techniques including IQR (Interquartile Range) Threshold, RMSE (Root Mean Squared Error) Threshold, Bollinger Bands, and Isolation Forests. We prove that the RMSE Threshold achieves the best precision with the lowest number of false fault alerts, while the IQR Threshold method provides the best recall by detecting the highest number of true fault alerts.