Applying Bi-LSTM Deep Learning Network for Power Consumption Anomaly Detection

In this paper, a Bi-directional Long Short Term Memory (Bi-LSTM) network is established for power consumption anomaly detection. The advantage of the Bi-LSTM is that it combines a forward LSTM and a backward LSTM, and can capture both forward and reverse temporal dependencies and periodicity characteristics in the power consumption data. To address the imbalance between normal and abnormal samples during model training, an oversampling strategy is employed to balance the power consumption dataset. After that, anomaly detection on a real-world power consumption dataset provided by the State Grid demonstrates that the proposed network achieves improvements of Precision (P), Recall (R), AUC and F1-score over classical CNN and LSTM, respectively. These results confirm the feasibility and superiority of the proposed network for anomaly detection in power consumption data.