Abstract:
Buildings consume 50% of the total available electrical energy. Studies show that up to 20% of the energy gets waste due to several reasons such as appliance misconfigurations, faults, and abnormal user behavior. The instances of abnormal energy consumption by various electrical appliances are known as anomalies, and it is important to detect such anomalies in a timely manner.
Several works use intrusive appliance-level monitors, i.e., submetering, for detecting
anomalous behavior of appliances. These approaches detect anomalies, but the intrusive monitoring approach is not scalable as it requires to have a separate monitor for each appliance in a home. A more practical approach is to use the household's aggregate consumption data from a single smart meter. Existing methods using smart meter data suffer from two limitations: (i) they result in a high number of false alarms which makes them unreliable, (ii) they only detect anomalies and do not identify the anomaly causing appliance, which makes it difficult for building administrator to take prompt action.
Using smart meter data, we handled the problem of anomaly detection both at coarse and _ne-granular levels. At coarse-level, we proposed a density-based anomaly detection approach namely, Monitor, which takes the power readings of several days as input and outputs anomaly score for each day consumption. At the granular level, we proposed an anomaly detection approach namely, RIMOR, and also evaluated the efficacy of Non-Intrusive Load Monitoring (NILM) for identifying anomalous appliance. Results show that proposed approaches reduce the false alarm rate considerably and also identify the anomalous appliances.
We did all our experiments on publicly available datasets namely Dataport, REFIT, AMPds, and ECO. These datasets have both aggregate and appliance-level energy consumption data, but none of them provides information about anomalous instances. We created and publicly released detailed anomaly annotations for REFIT at appliance-level and the remaining three datasets at the aggregate-level. Furthermore, we collected a context-rich four-year energy dataset from our campus and made it publicly available so that it can be leveraged to propose newer anomaly detection approaches with higher accuracy and take the research in a new direction.
