Time Series Anomaly Detection for Smart Grids via Multiple Self-Supervised Tasks Learning

Abstract

Smart grids have achieved tremendous success in energy infrastructure and also play an vital role in our daily lives. As an important auxiliary service, anomaly detection in smart grids, e.g., detecting abnormal activities and anomalous events from the electricity data, is very crucial to enhance the reliability and practicability for smart grids. In this paper, we mainly focus on the anomaly detection for time series data in smart grids. Although lots of anomaly detection approaches have been developed for time series data, most of them can not capture the mutual anomalous patterns dynamically among different sensors. In other words, dynamic correlations between different time series are always ignored, which consequently results in high false positive rates. To tackle the problem, we propose a self-supervised framework consisting of multiple tasks, i.e., inter-sample comtrastive task, intra-sample prediction task and reconstruction task, without using groundtruth labels. Technically, in the first task, we construct contrastive pairs from different samples (inter-sample pairs) and propose to use the multi-headed attention model to capture the dynamic mutual anomalous patterns. Furthermore, the prediction task with multiple fully connected (MLP) layers stacked with dropouts is proposed to capture the temporal intra-sample trends in time series data. In the final task, we use Gate Recurrent Unit (GRU) based autoencoders to compute the reconstruction error. The final loss is combined with inter-sample contrastive loss, reconstruction loss and intra-sample prediction loss. Compared with several baselines on a real-world dataset, the experimental results demonstrate that our proposed approach can effectively detect anomalies and significantly outperforms the state-of-the-art.