Anomaly Detection in Electricity Consumption Data using Deep Learning

Anomaly detection in electricity consumption data is one of the most important methods to identify anomalous events in buildings and electric assets, such as energy theft, metering defect, cyber attacks and technical losses. In this paper, a novel deep learning based approach is presented to detect anomalies in electricity consumption data one hour ahead of time. We address this challenge in two stages. First, we build an Long Short-Term Memory (LSTM) based neural network model to predict the next hour sample. Second, we use another LSTM autoencoder to learn the features of normal consumption. The output of the first stage is used as an input to the LSTM autoencoder. The LSTM autoencoder will learn the features of normal consumption and the input will be similar to output when applied. For anomalies, the input and output will be significantly different. The Exponential Moving Average (EMA) is used as a threshold and two types of anomalies are distinguished, local and global anomalies. Several weather features are considered in this study, such as pressure, cloud cover, humidity, temperature, wind direction and wind speed in addition to temporal and lag features. A feature selection method to find the optimal features that achieve good results is also implemented. We validate the proposed approach by comparing the detected anomalous consumption and the normal consumption within the same period, and the results demonstrate a drastic increase in electricity consumption during the anomalous periods. The results also show that the temporal and lag features have improved the efficiency and performance of the proposed method.