Abstract:With the continued development of new distribution systems in China, distribution networks are being transformed into active bidirectional systems characterized by flexibility, digitalization, and intelligence. However, multidimensional risks are being introduced by the large-scale integration of distributed resources and increasingly open communication environments, including equipment degradation, component outages and operating-limit violations; measurement sparsity, unknown topology and false data injection (FDI) attacks; and communication congestion, link failures, packet loss and denial-of-service/distributed denial-of-service (DoS/DDoS) attacks. These risks are deeply coupled through energy and information flows, through which complex cross-domain cascading failures can be triggered. To address such multilayer coupled risks, a three-layer physical-digital-information architecture is established, and recent applications of graph neural networks (GNNs) in risk management and control for new distribution systems are systematically reviewed after the risk evolution mechanisms at each layer are analyzed. The relationships among multidimensional risk characteristics, specific graph learning tasks, and different GNN architectures are identified and summarized, while particular emphasis is placed on the advantages of GNNs in capturing non-Euclidean topological dependencies and addressing cross-layer coupled risks. Finally, the limitations of existing data-driven methods are analyzed, and future research directions are outlined in terms of physics-guided modeling, cross-layer multimodal graph fusion, and cloud-edge collaborative computing. A reference for the secure operation of highly resilient new distribution systems is expected to be provided through this review.