A Bayesian optimized graph attention network (BO-GAT) based power flow calculation method is proposed for distribution networks. This method addresses the low computational speed and reliance on complete line parameters of conventional power flow methods. It also overcomes the limitations of existing data-driven approaches in handling frequent topology changes. The method utilizes the topology and node features of the distribution network to construct graph data, and calculates attention coefficients using the graph attention mechanism. By capturing correlations between nodes, the method enhances the adaptability of the power flow regression model to topology changes. The Bayesian optimization (BO) algorithm is introduced to optimize the hyperparameters, further enhancing the performance of the model. The model's regression accuracy and computational efficiency are evaluated on the improved IEEE 33-node system. The results demonstrate that the proposed method can achieve rapid power flow calculation without specific line parameters. It also exhibits strong robustness and topology generalization capability under measurement information loss and topology changes. Moreover, even with a significant increase in wind and solar energy penetration, the calculation accuracy remains high. Finally, the applicability of the proposed method to large-scale distribution networks is further validated on the IEEE 141-node system.