Scalable Prediction-based Online Anomaly Detection for Smart Meter Data

Abstract Today smart meters are widely used in the energy sector to record energy consumption in real time. Large amounts of smart meter data have been accumulated and used for diverse analysis purposes. Anomaly detection raises the big data problem, namely the detection of abnormal events or unusual consumption behaviors. However, there is a lack of appropriate online systems that can handle anomaly detection for large-scale smart meter data effectively and efficiently. This paper proposes a lambda system for detecting anomalous consumption patterns, aiming at assisting decision makings for smart energy management. The proposed system uses a prediction-based detection method, combined with a novel lambda architecture for iterative model updates and real-time anomaly detection. This paper evaluates the system using a real-world data set and a large synthetic data set, and compares with three baselines. The results show that the proposed system has good scalability, and has a competitive advantage over others in anomaly detection.