Abstract: Conventional optimal power flow methods encounter such problems as high computational overhead and insufficient real-T response capability when applied to large-scale power systems under complex operation scenarios. To address this issue, this paper proposes a physically guided globally topology-aware graph neural network based on a state-conditional continuous edge-direction message passing mechanism. In the proposed method, continuous direction angles are dynamically generated based on node states and line impedances, and adaptive regulation of symmetric diffusion and asymmetric propagation is realized via anti-symmetric constraints and direction confidence residual gates. Furthermore, a unified modeling framework is constructed that integrates direction gates, multi-hop propagation, global attention and power flow physical loss. Experiments conducted on the IEEE-39-bus, 118-bus and 300-bus systems demonstrate that compared with the GAT baseline, the proposed model reduces the L2 loss from 0.042, 0.073 and 0.198 to 0.011, 0.015 and 0.042 respectively, while maintaining millisecond-level inference efficiency and sound prediction stability under N-k topology variation scenarios. The results of ablation experiments and direction visualization verify the effectiveness of the proposed model mechanism in characterizing dynamic asymmetric electrical coupling and enhancing the interpretability of message propagation. The experimental results indicate that within the scope of the tests conducted in this paper, the proposed method achieves a balanced trade-off among prediction accuracy, inference efficiency, adaptability to topology variation and interpretability of message propagation while maintaining low inference overhead.