A fast reliability assessment method for power system using self-supervised learning and feature reconstruction

Abstract

Under the global zero-carbon campaign (United Nations Climate Change, 2021), more stochastic renewable generations are being integrated into the system. This trend significantly expands the state space and increases the computational burden for the model-driven reliability assessment. Data-driven approaches are developed to improve efficiency based on artificial intelligence. However, the requirement of large-scale samples limits it in applications. To address that, this paper adopts a self-supervised stage to avoid the high cost of labeling, while ensuring the efficiency and accuracy of reliability assessment. The training process of this method is split into two stages. In the first stage, feature reconstruction and unsupervised learning are used to provide the initial network parameters. Thereafter, the second learning stage can be trained in a task-agnostic way with fewer labels. The results of case study demonstrate the effectiveness of the proposed approach.

Keywords

• Reliability assessment
• Self-supervised learning
• Feature reconstruction
• Supervised learning
• Data drive