Under high-dimensional and nonlinear stochastic power system environment, artificial intelligence (AI) is becoming a promising alternative towards the urgent demand of real-time AC optimal power flow (OPF). However, traditional AI only fit the settlement results of other method, and is unable to find in-depth law of power flow. This may result in an undesired poor generalizability. To address this issue, a unsupervised OPF model combining physics- guided graph neutral networks and Lagrangian dual is proposed. The model incorporates the physical constraints into training tensor graph, as well as a dual Lagrangian method to satisfy the physical and engineering constraints present in the OPF. At the same time, since topology of power system may vary, the branch characteristics are embeded into node graph to adapt to grid topology contingency. Numerical tests on benchmarks demonstrate that, in unseen operating condition, the proposed method enables a near or even better solution than conventional optimization algorithm, but consumes much more than 100 times calculation efficiency.