This study presents a reduced-order model (ROM) of a large-scale transcritical CO2 heat pump designed by MAN Energy Solutions, currently operating in Esbjerg and under construction in Aalborg, Denmark. The original high-fidelity model, implemented in Modelon Impact and presented at Modelon Innovate, captures the complex thermodynamic and control dynamics of the heat pump but comes with high computational cost, making it impractical for real-time applications. To address this issue, we trained a neural network ROM using simulation data from the Modelica-based model, enabling significant reduction in simulation time without compromising accuracy. We demonstrate how the ROM is integrated back into the Modelica framework to facilitate fast simulations, paving the way for real-time control scenarios and digital twin implementations. Performance evaluations show that the ROM retains key dynamic behavior and delivers simulation speed-ups by more than two orders of magnitude. The model captures start-up and shut-down scenarios, assessing system reliability and transient behavior, and evaluates the heat pump’s flexibility for grid balancing, responding to variable load demands and grid needs. This approach not only accelerates system-level studies but also supports online applications such as fault detection, predictive control, and system optimization. Our work highlights the potential of combining physics-based modeling with data-based machine learning techniques to bridge the gap between simulation fidelity and computational efficiency, crucial for the development of next-generation digital twins in the energy domain.