A physics-guided RNN-KAN for multi-step prediction of heat pump operation states

Dynamic modeling and operating state prediction for heat pump components are crucial for achieving energy-efficient heating and form the foundation for advanced optimization and control strategies. Heat pump modeling faces two primary challenges: accounting for inherent thermal hysteresis and capturing the nonlinear relationships between model parameters. This paper presents a physics-guided, component-level dynamic modeling framework for heat pumps. Each component is modeled using an innovative RNN-KAN, with input and output variables grounded in physical principles. By sequentially coupling component models based on the thermodynamic cycle, the system achieves multi-step state predictions through a recursive process. To balance prediction accuracy and computational efficiency, a prediction-error-based training method is introduced. This study pioneers the application of the KAN combined with RNN in heat pumps, providing a novel dynamic modeling method and broadening the scope of application. The results demonstrate that the RNN-KAN model effectively captures nonlinear relationships and accounts for thermal hysteresis, improving heating capacity prediction accuracy by 7.75 % over conventional RNN models. Furthermore, the error-based training method enhances prediction accuracy by 4.38 % compared to batch training while reducing computational time by 53.7 % relative to online training methods.