Anomaly detection in smart grid using a trace-based graph deep learning model

Electricity plays a significant role in the everyday lives of people. Researchers have long been interested in the classification problem of electric power anomaly detection. Anomaly detection can stop little issues from snowballing into unmanageable issues. In addition, it helps cut down on energy waste. Existing anomaly detection models mostly ignore the spatial attribute of electricity consumption data. They would primarily emphasize the time series information contained within the energy consumption data. Furthermore, the trace has the ability to precisely reconstruct consumer pathways; the smart grid can thus use it to detect anomalies. To fill this research gap, we propose a trace-based graph deep learning model to detect anomalous consumers in the smart grid. An unsupervised encoder-decoder is used in the proposed model. First, our model combines traces using an efficient unified graph representation and provides quality scores. Then, the long short-term memory network extracts the temporal attributes, while the graph neural network extracts the spatial attributes. Finally, it computes the anomaly score by adding two hyper-parameters with two-part loss values. We conducted experiments on power consumption data that was gathered from an open-source dataset. The proposed model performs better than a range of standard anomaly detection models. The F1-score of our model is 94.60%, and the AUC is 98.90%. Experiments show that our model is stable even in extreme data imbalance.