Liquid neural network-driven intelligent topology perception and path anomaly diagnosis for Ipv6 smart grids

The deployment of Segment Routing IPv6 (SRv6) in smart grids introduces highly dynamic topologies and continuous-time telemetry streams. Traditional discrete-time Recurrent Neural Networks (RNNs) struggle to process these asynchronous, irregularly sampled data streams without suffering from severe discretization information loss, limiting their real-time diagnostic capabilities. To resolve this, this paper proposes an intelligent topology perception and path anomaly diagnosis framework driven by Liquid Neural Networks (LNNs). The novelty of this work lies in formulating network state transitions as continuous-time dynamics governed by Ordinary Differential Equations (ODEs). This approach natively accommodates irregularly sampled telemetry data without discretization, allowing the network’s internal states to dynamically adapt to transient anomalies. Furthermore, we design a multi-task learning architecture that unifies real-time topology inference and probabilistic fault forecasting into a single end-to-end framework. Experimental evaluations on a simulated IPv6 smart grid demonstrate the objective superiority of the proposed model over state-of-the-art baselines. Specifically, the LNN framework achieves a 98.0% F1-score in topology awareness (outperforming Informer by 2.5%), reduces the False Positive Rate (FPR) in anomaly detection to 4.7%, and provides an Early Warning Time (EWT) of 8.5 s, which is a 13.3% improvement over the best baseline model. Comprehensive ablation and scalability studies further validate the computational feasibility and robustness of the LNN framework under high data-missing rates, providing a highly reliable monitoring mechanism for mission-critical power networks.