Enhancing anomaly detection methods for energy time series using latent space data representations

The increasing number of recorded energy time series calls for an automated operation of smart grid applications such as load forecasting and load management. While these applications require anomaly-free data to perform well, the recorded data often contains anomalies. The numerous methods to detect these anomalies are typically applied directly to the recorded data. However, for other tasks such as forecasting, promising performance has been achieved when directly applying methods to a meaningful feature space of the data, i.e., the latent space data representation. We, therefore, propose a novel approach to generally enhance anomaly detection methods for energy time series by taking advantage of their latent space representation. We create latent space data representations using a conditional Invertible Neural Network (cINN) and a conditional Variational Autoencoder (cVAE) and directly apply existing supervised and unsupervised detection methods to this representation. We evaluate the latent space data representation qualitatively by visualizing the separation of anomalies and non-anomalous data. We also quantitatively evaluate our approach by applying supervised and unsupervised detection methods to real-world load data containing two groups of artificially inserted anomalies: technical faults and unusual consumption. We show that our approach generally improves the anomaly detection performance of the considered methods while only moderately increasing computational cost.