Accurate control of chiller systems is critical for energy efficiency, yet conventional data-driven models often fail under unseen conditions due to poor generalization and unquantified uncertainty. This study proposes a hybrid AI chiller model using transfer learning (TL), combining physics-based knowledge with measured data, and applies it to a stochastic optimal control (SOC) framework. Uncertainty is estimated using a Monte Carlo simulation, allowing risk-aware control decisions. The SOC algorithm minimizes total power consumption by optimizing control variables while making uncertainty-aware decisions that reduce the risk of overcontrol. The hybrid AI model could achieve energy saving by 30.5% and decrease the degree of lower uncertainty compared to a baseline ANN model. These results highlight the value of incorporating both physical consistency and uncertainty quantification in AI-based control, enabling more robust and reliable HVAC operation under real-world variability.
Kim et al. (Wed,) studied this question.