Unsupervised Learning for Online Abnormality Detection in Smart Meter Data

The analysis of abnormalities in smart meter data has applications in load forecasting, cyber security, fault detection, electricity theft detection, demand response, etc. Abnormality is broadly defined in this paper as any unusual electricity consumption instance or trend that falls outside of the normal usage patterns for each load, whether in terms of magnitude, time of usage, etc. Unusual electricity consumption can have different signatures and different duration of time. This paper aims to evaluate the performance of four unsupervised machine learning methods for abnormality detection on real-world smart meter data, namely prediction-based regression, prediction-based neural network, clustered-based, and projection-based methods. Different types of features, such as load-based, contextual, and environmental, are investigated to construct the data-driven models. It is shown that different abnormality detection methods have different ability for detecting different types of abnormalities; and their performance depends on the set of features used for training the method. Accordingly, different types of features are scrutinized for each abnormality detection method.