The traditional approach based on human experience has limited diversity in operating scenarios, making it difficult to fully characterize the complex operating boundaries of high proportion new energy power systems. Operation model arranging based on this approach may result in safety risk blind spots in the system. This article proposes a typical operation mode extraction method based on convolutional neural networks and self-attention mechanisms. Firstly, an autoencoder model is constructed using convolutional neural networks to intelligently extract complex spatial coupling relationships between variables in the operation of high proportion new energy grids. Secondly, based on the extracted operational characteristics of the power grid, a feature clustering layer is introduced, which is jointly optimized with an autoencoder model to achieve clustering. Then, the clustering results of the operation mode are characterized by the proposed indicators of the new energy-load-traditional energy combination mode and the clustering effect evaluation indicators. Finally, the sample set represented is expanded by class centers and static security assessments are conducted. The calculation results show that this method can effectively explore the spatial correlation and complex combination patterns of high-dimensional operating variables in the power grid. The safety verification carried out on this basis helps to differentiate and characterize the safety risks of high proportion new energy power systems between different modes, providing strong support for the formulation of typical operating modes of new power systems.