A method for chiller performance modeling via SKR-based neural network under physical constraints

The chiller is a critical component in heating, ventilation, and air conditioning (HVAC) systems, and its operating efficiency has a major impact on the overall energy performance of buildings. As a key performance indicator, accurately modeling the coefficient of performance (COP) of the chiller is critical. However, COP is significantly affected by system dynamics and uncertainties, which limit traditional neural network-based modeling methods due to their neglect of temporal dependencies in state evolution. To solve this problem, this paper proposes an SKR-based neural network (SKRNN) for COP modeling based on stable kernel representation framework. Firstly, a feature selection strategy is designed, which integrates minimum redundancy maximum relevance and autocorrelation/cross-lagged correlation analysis to distinguish between input and state features and eliminate redundancies. Secondly, an SKRNN modeling architecture is developed, which explicitly characterizes the nonlinear dynamic process of COP as a function of internal and external factors via an observer. Finally, a physically informed constraint loss function based on thermodynamic mechanisms is introduced to enhance the modeling performance of the model. SKRNN combines the strengths of state-space modeling and nonlinear fitting of neural networks, and achieves dynamic real-time COP modeling through output feedback. Its effectiveness has been validated with operational data from an actual building chiller system, comparative experiments with mainstream modeling methods demonstrate that the proposed method can effectively improve COP modeling accuracy.