<p>Direct current microgrids (DCMGs) are highly susceptible to post-fault voltage instability due to the absence of inherent inertia, making reliable stability assessment essential. This paper proposes a hybrid framework integrating dwarf mongoose optimization (DMO) with an attributed multi-order graph convolutional network (AMGCN) for binary post-fault voltage stability classification. The Electrical Grid Stability Simulated Dataset (EGSSD) is first preprocessed using a generalized multi-kernel maximum correntropy Kalman filter (GMKMCKF) to suppress noise and enhance signal quality. The honey badger algorithm (HBA) is then used to select significant stability-related features. The selected features are represented as nodes in a graph structure, where inter-feature relationships are modeled through correlation-based adjacency matrices. The AMGCN conducts supervised learning by cross-entropy loss to determine whether the operating conditions are stable or unstable, whereas the DMO optimizes the key hyperparameters to increase the level of generalization. The findings indicate that the suggested framework DMO–AMGCN achieves classification accuracies of 99.2 and 98.9% in stable and unstable conditions, respectively, with a mean squared error of 0.0084 which is better in relation to conventional artificial neural network (ANN), deep neural network (DNN), and other deep learning (DL) models. The results prove the efficiency and computational feasibility of the suggested method as a reliable monitoring of voltage stability in the DCMG environments.</p>