A Graph-Temporal Joint Modeling Approach for Communication Failure Risk Prediction in Distribution Networks

Communication failures in cyber-physical distribution networks (CPS-DNs) can lead to mis-operation of controllable devices, resulting in voltage violations, power flow imbalance, and potentially cascading faults. To address the challenge of risk prediction under uncertain communication conditions, this paper proposes a graph-time collaborative learning framework that integrates graph representation learning with temporal sequence modeling. Specifically, the network state at each time step is modeled as a dynamic graph, where node features encode voltage, power, communication status, and device type. A Graph Attention Network (GAT) is used to adaptively aggregate neighborhood information, capturing topological heterogeneity and the impact of local communication faults. The resulting sequence of graph embeddings is then fed into a Long Short-Term Memory (LSTM) network to model the timedependent evolution of risk. A simulation case based on an enhanced IEEE-33 node system demonstrates the framework’s ability to predict system-level risk indices under various fault scenarios. Extensive ablation experiments validate the effectiveness of both the graph-based and temporal components. The proposed approach provides a robust and scalable solution for proactive risk assessment in smart distribution networks.