Anomaly Detection Model for Data Interactions Based on Temporal Features in Electric Power Data Centers

The current power data centers face the risk of data leakage during the data exchange process. Given the increasing scale and complexity of power data centers, traditional methods for identifying anomalies in the temporal characteristics of data interactions are not applicable. However, there is currently no specific model available for identifying abnormal data interaction behavior in the context of power data centers. Directly using existing models would result in low efficiency and fail to meet the requirements of power data center systems. This article proposes a model, VAE-GAN, for identifying abnormal data interaction behavior with multivariate temporal characteristics specifically designed for power data center business scenarios. The Variational Auto Encoder (VAE) architecture allows the model to learn in an unsupervised manner. The use of Generative Adversarial Networks (GAN) and their architecture helps address the drawbacks of VAE networks in terms of smoothness and reconstruction ability in latent spaces, thereby improving the model's generation and reconstruction performance. Experiments were conducted on a dataset provided by a power company to study the effectiveness of the VAE-GAN method in power data center business scenarios. The results demonstrate its usability in the context of power data centers.