Anomaly detection of building systems using energy demand frequency domain analysis

This paper presents and demonstrates a method to quickly identify when regular periodic activities, such as a daily night setback on a thermostat, are inappropriately configured or accidentally reset. Anomalies in periodic building operations are identified by analyzing smart meter electrical demand data in the frequency domain with a weekly travelling time window instead of using time domain functions such as load factor. Initial experiments on a real site found that spectral energy signals for periodic (frequency) hours of 4, 6, 8, 12 and days 1, 3.5 and 7 to be greatly reduced when a device is not functioning appropriately. In addition, the ratio of the DC offset (0 Hz) energy with the other higher periodic energies can normalize the periodic energies to a relative index that can then be used for comparing other seasons and other buildings for periodical performance.