A Graph Recurrent Attention Network for Multivariate Time Series Anomaly Detection

In this paper, we propose a Graph Recurrent Attention Network (GRAN) for multivariate time series anomaly detection. GRAN combines Graph Attention Networks (GATs) and Long Short Term Memory (LSTM) to learn both the temporal information of variables and the relationship between them. In addition, we have improved the self-attention mechanism in GAT to better allocate the weights of the relationship between variables based on the time series information. Experiments on commonly used public datasets show that our method outperforms baseline, and the effectiveness of our design is demonstrated by ablation experiments. This method has important application value for the realization of abnormal early warning and real-time monitoring of oil and gas stations, and provides strong support for ensuring the stability and safety of oil and gas production.