Hybrid AI chiller model using transfer learning: Aleatoric and epistemic uncertainty

This study proposed a hybrid AI chiller model using transfer learning (TL) to improve prediction accuracy and extrapolation reliability beyond the measured operating conditions. The model integrates physics-based knowledge from chiller specifications with measured operational data, thereby enabling physically consistent predictions while leveraging data-driven insights. An artificial neural network (ANN) was pre-trained with synthetic data generated from an empirical chiller model and fine-tuned using measured data from five chillers. Model accuracy was assessed using the coefficient of variation of the root mean square error, whereas uncertainty was quantified using Bayesian deep learning with Monte Carlo dropout to evaluate the aleatoric (AU) and epistemic uncertainties (EU). The results demonstrated that both the ANN and TL models achieved acceptable accuracy within the ASHRAE guideline limit of 30%. However, the TL model exhibited significantly lower uncertainty, with the AU, EU, and total uncertainty (TU) reduced by 73.4%, 70.4%, and 72.2%, respectively. The reduced AU was attributed to the pre-training on refined specification data, whereas the lower EU resulted from physics-based knowledge that improved extrapolation in regions with limited data. Overall, the proposed TL-based hybrid AI model offers a practical and reliable solution for chiller performance prediction, supporting energy-efficient system operation and enhancing reliability under unmeasured operating conditions.