<p>Accurate device-type identification (DI) is critical for ensuring the security and stability of large-scale Internet of Things (IoT) deployments in smart grids. However, existing traffic-based DI methods often struggle in dynamic environments, as they fail to capture the temporal evolution of device behaviors, overlook complex inter-device dependencies, and lack robustness to the sparse or incomplete data common in practice. To address these challenges, we propose DRID, a novel Spatiotemporal Dynamic Relational Framework for Robust IoT Device Identification. DRID jointly captures structural communication patterns and multi-scale temporal dynamics via a structure–time interaction mechanism and multi-scale temporal modeling, while leveraging a differentiation-aware adaptive learning strategy to selectively enhance discriminative features under sparse or noisy traffic conditions. Extensive evaluations on two public IoT traffic datasets demonstrate that DRID consistently outperforms state-of-the-art baselines across diverse sampling scenarios. By effectively fusing structural and temporal information while maintaining robustness under data scarcity, DRID provides a scalable and accurate solution for IoT device identification, advancing secure and intelligent management of critical smart grid infrastructures.</p>