The deployment of sensors, smart terminals, and edge gateways is driving the transformation of traditional power grids into the power Internet of Things (IoT). However, asset operation and maintenance still face data fragmentation, insufficient topology-aware modeling, delayed risk response, and high operational costs. This paper proposes a Digital Twin-assisted Federated Graph Learning framework (DT-FGL), which constructs a layered architecture consisting of physical assets, data communication, hybrid twin modeling, federated graph risk prediction, and decision services. The framework integrates electrical topology, device status, environmental conditions, communication quality, and maintenance records to generate a heterogeneous asset graph. Here, rare-fault recall refers to the recall of the minority positive class, defined as equipment faults occurring within the future prediction horizon. Cascading risk is modeled through learned multi-hop graph propagation over electrical, spatial, communication, co-failure, and resource-sharing relations, rather than through full physical power-flow simulation. Node-level risk probabilities are then mapped into inspection and maintenance decisions through a budget- and risk-constrained optimization layer. Synthetic benchmark experiments show that DT-FGL improves rare-fault recall to 0.317 and achieves an AUC of 0.804, while reducing operation and maintenance costs by 43.8% relative to fixed-interval maintenance. Because the current validation is based on synthetic benchmark data, the reported gains should be interpreted as simulation evidence that requires further verification on field power IoT datasets.
