Abstract: With the massive integration of new energy sources, the randomness of the power system has increased significantly, and the data distribution of the operation scenarios is scattered and uneven, which reduces the applicability of the existing scenario generation methods. Traditional scenario generation based on human experience and models pays more attention to some characteristics of the scenario and ignores the data distribution of the scenario. Data-driven scenario generation methods focus on describing the data distribution of the scenario, while some operation scenarios with low probability and high risk are easily overlooked. To address this issue, this research paper presents a novel approach for scenario generation, combining graph representation learning and feature-guided operation scenario generation based on the graph representation of grid operation scenarios. The proposed method mines grid operation features and incorporates the desired features into the model through data and knowledge fusion, ensuring feature-guided operation scenario generation while maintaining and preserving the distributional characteristics as far as possible. Ultimately, the verification results obtained from a system that incorporates a substantial amount of renewable energy sources and considers the operational risks of the power grid demonstrate that the proposed model outperforms the traditional scenario generation method. The proposed model not only enhances the generation efficiency of specific operational scenarios, but also ensures the consistency and diversity of the generated scenarios. Moreover, this paper provides comprehensive data support for machine-aided decision-making in power system dispatching.