This study explores the development of data-driven surrogate models for Seasonal Thermal Energy Storage (sTES) suitable for integration into Model Predictive Control (MPC) frameworks. Using measurement data from the Dronninglund pit storage system, a validated Modelica model serves as the foundation for generating training data for various surrogate modeling approaches. The objective is to assess whether data-driven models can provide a computationally efficient alternative to physics-based simulations while maintaining sufficient accuracy for control applications. The study focuses on the surrogate model performance and its ability to replicate temperature stratification, heat losses, and ground temperature evolution. Multiple machine learning algorithms are evaluated based on predictive accuracy, generalization capability, and computational efficiency. The study also examines how different model structures and training strategies influence performance. Additionally, the feasibility of integrating the surrogate model within an MPC setup is analyzed, considering both computational requirements and control-relevant accuracy. By systematically comparing surrogate models against the original Modelica-based approach, the study provides insights into the trade-offs between complexity, accuracy, and real-time applicability. The results contribute to the broader discussion on optimizing sTES operation through data-driven modeling, offering guidance on selecting appropriate methods for predictive control in large-scale thermal energy storage systems.