Multi-output model of medium-temperature chillers for digital twins: A comparative study of steady-state and dynamic modeling approaches

Chiller modeling is essential for ensuring efficient chiller operation. The existing chiller models are mostly single-output steady-state models that cannot accurately capture the dynamic behavior of chillers and cannot meet the needs of digital twins. In this work, a multi-output model framework was proposed to facilitate the development of a digital twin chiller. Subsequently, three steady-state and three dynamic chiller models were developed based on a medium-temperature case. The hyperparameters of the six candidate models were optimized. To systematically evaluate model suitability, we introduced two novel metrics: the univariate error, which quantifies prediction accuracy for individual variables, and the model overall error, which aggregates errors across all variables to assess comprehensive performance. A comparative analysis was then conducted to contrast the best steady-state and dynamic models, evaluating their overall error and dynamic responsiveness. The study results show that: The chiller power consumption of all models exhibit the lowest prediction accuracy, followed by evaporator outlet water temperature and condenser outlet water temperature. The support vector regression (SVR) model is the best of the steady state models with model overall error of 10.84 %, and the gate recurrent unit (GRU) model is the best of the steady state models with model overall error of 3.67 %. Notably, the GRU model demonstrates superior accuracy in predicting evaporator outlet temperature($T_{\mathrm{eo}}$), condenser outlet temperature($T_{\mathrm{co}})$ and chiller power consumption($P)$ and better captured transient fluctuations in these variables during chiller start-up and load changes compared with the SVR model. The findings provide a methodological foundation for developing digital twin models and optimizing intelligent operation/maintenance strategies for chillers.