Traditional smart distribution network state estimation faces some problems, such as out-of-sync sampling, missing measurements, topology changes, and heterogeneous data formats of SCADA, PMU, smart meters, feeder terminals and environmental sensors. This paper proposes a state estimation model for multi-source data fusion, which unifies asynchronous measurements in the framework of graph-aware computing. Buses, feeders, and transformers are represented as topological embeddings, and irregular measurements are aligned through a time-fused encoder weighted by sources-level confidence. The layer output node voltage, phase Angle, branch current, active and reactive power states are estimated, and the calibration component identifies abnormal measurements and accounts for confidence changes. The model is tested on the IEEE 33-node distribution network simulation dataset, which contains 52,800 samples, and the provincial distribution network dataset, which contains 286,000 operating records. Experimental results show that in the actual online deployment scenario, the proposed model reduces the voltage MAE to 0.0136 p.u., the state recognition accuracy reaches 96.2%, and the average inference delay is 18.7 ms.
