Abstract: As digital power grids advance towards system-level intelligence, the multi-layered synchronization and collaboration spanning "IoT-Data-Graph Models" constitutes a core metric of maturity. Addressing the limitations of conventional approaches in effectively characterizing and identifying such synchronization, this paper reframes the "Four Synchronizations" maturity assessment as a spatiotemporal synchronization state recognition problem. A multi-task learning model based on spatiotemporal graph neural networks is proposed. The model features a four-tier architecture that processes multi-modal inputs, represents dynamic heterogeneous graphs, and employs cross-layer attention mechanisms to concurrently output maturity evaluation, synchronization diagnostics, and bottleneck analysis. Empirical validation using provincial power grid data demonstrates that the proposed method significantly outperforms traditional weighted approaches across all three tasks, while exhibiting superior assessment stability and engineering applicability.