Extreme outage prediction in power systems using a new deep generative Informer model

Abstract

Extreme weather events have made growing concerns over electric power grid infrastructure as well as the residents living in disaster areas. Moreover, the potential damages due to the extreme events can make serious challenges for supply reliability and security, leading to widespread power outages in power systems. This paper proposes a deep learning-based framework for power data rebalancing and outage prediction in power systems to cope with the extreme events. To this end, we propose an Adaptive Wasserstein Conditional Generative Adversarial Network for data generation. Also, we propose a new Wasserstein Bidirectional Generative Adversarial Network with the Informer model, embedded in both the Generator and Discriminator Networks, plus an Encoder Network for the outage prediction in power systems. Two-step classification approach has been used in the proposed outage prediction model: classifying the power grid components into impacted and non-impacted categories and classifying the impacted category into in-service and out-of-service categories. In addition, a new classification-specific loss function is proposed for the minimax objective function of the Vanilla Generative Adversarial Network to improve the prediction performance in the latent space. Evaluation results of the proposed model and 15 comparative models in three groups using six evaluation metrics on a real-world test case demonstrate the superiority of the proposed model compared to all comparative models. These results confirm that the proposed outage prediction model can be effectively employed for accurately predicting extreme outages in power systems.