This paper presents optimal control strategies of a cooling system for an existing building comprising several dozen chillers, pumps, and cooling towers. Particularly, such a large cooling system is designed to respond to highly time-varying cooling loads and guarantee a stable indoor environment. However, the existing control follows a predetermined rule-set for determining the number of operating devices which is presumably not the optimal operation for the facility. In this study, we introduced a physics-informed data-driven simulation model for the chillers that combines our domain knowledge of chiller dynamics with measured data. In other words, the model uses both operational datasets and the chiller’s specification information as a training dataset to address the chillers’ actual dynamics. The chiller model was developed using a Gaussian process emulator, one of the machine learning methods and can predict real-time time-varying coefficient of performance (COP). Subsequently, we applied Model Predictive Control (MPC) to the cooling system in the target building. MPC is designed to determine the optimal number of operating devices minimizing energy use by the system. The proposed control, MPC could save energy by 5–10% compared to the existing control, or rule-based control. An additional noteworthy benefit of the physics-informed data-driven model is that it can efficiently predict the chiller’s operating performance beyond the trained data because it contains physical knowledge. In other words, it exhibits better extrapolation behavior than a purely data driven model.