Abstract: Typical operation modes are important for power system planning and dispatching. Due to the high-dimensional, non-linear, and uncertain characteristics of power system, the generation of operation modes often faces combinatorial explosion problems, which makes it challenging and inefficient for operators to generate specified types of operation mode in needs, such as operation modes which has a low security margin of selected flowgate. To this end, this paper proposed a sample generation framework and training method combing generative adversarial network (GAN) and model-based transfer learning, which could efficiently obtain a high-performance typical operation mode sample generation model with a small amount of data and fine-tuning processes. Firstly, a GAN model oriented to the generation of operation modes was designed, and the common characteristics of different types of operation modes were fully learned through the basic model. Besides, a training method suitable for GAN with model-based transfer learning was proposed, and the target model after fine-tuning from the basic model could efficiently and accurately generate a large amount of typical operation samples. The proposed method has been verified on New England 10-machine 39-bus system with safety margin of flowgate as target. The result shows that proposed method could generate operation modes with given safety margins of flowgates, which could solve the problems of inadequate typical operation modes and provide sufficient data for subsequent operation mode analysis.