This paper investigates data-driven decision-making methods for Security-Constrained Unit Commitment (SCUC) in power systems. Traditional physics-model-driven SCUC (PMD-SCUC) suffers from limited adaptability and efficiency in complex decision-making scenarios. To overcome these issues, a data-driven SCUC (DD-SCUC) framework is proposed. First, a hybrid deep learning model combining data-driven mapping and physical constraints is developed to predict unit outputs from load profiles, with clustering preprocessing to enhance accuracy. Second, an improved E-Seq2Seq architecture with multiple Encoder-Decoder networks is introduced to process elastic multi-sequence data, improving adaptability and precision. Furthermore, a composite decision-making method is designed for typhoon scenarios, integrating source-grid-load coordination based on game-theoretic optimization to enhance system resilience. Simulation results based on real-world grid data verify the feasibility and effectiveness of the proposed methods in improving decision accuracy, computational efficiency, and operational robustness.