Abstract: Power grid simulation is of great significance to power system planning, operating, and control decision- making. It is an important step in simulation calculation to analyze the stability of power system and the dominant instability mode (DIM) according to the massive simulation data, so as to provide support for the subsequent formulation of emergency control decision tables. In this paper, deep learning was used to overcome the problem that traditional methods are difficult to effectively distinguish rotor angle instability and voltage instability in the actual power grid. In order to reduce the dependence of the deep neural networks on labeled samples, this paper proposed a semi-supervised learning framework based on mean teacher (MT) with the virtual adversarial training (VAT) model for the intelligent identification of DIM in simulation analysis. The VAT-MT model constructed a teacher network and a student network respectively. The model training was enhanced by applying small disturbance to the features of all samples, and then input into the two networks to calculate the consistency loss. At the same time, the maximum disturbance direction was calculated by VAT to improve the performance of the model. Case studies were conducted on the China Electric Power Research Institute 36-bus system and Northeast China Power Grid. The results show that the proposed method can effectively reduce the labeling cost, and has the ability to adapt to the actualpower grid.