State estimation is the foundation of situational awareness in distribution systems and also a prerequisite for secure

reliable

and efficient operation. However

timely acquisition of the actual operating topology is hindered by limited measurement coverage and the lack of switch status information in real time. Meanwhile

as distributed energy resources and flexible loads continue to penetrate

increasingly diverse operating modes are observed

and topological uncertainty is further exacerbated by frequent switching

posing severe challenges to state estimation. To address these issues

this paper proposes a dynamic spatiotemporal graph neural network method with physical consistency for topology detection and state estimation. A design with a single encoder and multiple decoders is adopted

under which spatiotemporal features from multiple sources are extracted by a shared encoder and topology detection results and system states are jointly produced by the decoders. An adaptive dynamic graph mechanism is introduced

by which electrical correlations among nodes are inferred from measurements in real time

thereby addressing the difficulty of constructing a dynamic graph when only part of the topology is observable. In addition

power flow equations are embedded in the loss as physical consistency constraints. As a result

interpretability is improved and robustness across operating conditions is enhanced. As shown by simulation studies

high accuracy is maintained under sparse measurements and strong noise

and comparable accuracy is sustained for topologies unseen during training without transfer learning. Generalization across scenarios is thereby demonstrated

and a physically interpretable

dynamically adaptive solution for topology detection and state estimation is also provided.