Control-oriented modeling is critical for optimizing the practical operation of variable refrigerant flow (VRF) systems, yet varying topologies and non-linear coupling characteristics often cause purely data-driven models to struggle with generalization. Therefore, this work proposes a physics-guided data-driven performance map (PGDPM) method, which employs a strategy of component decoupling and system integration to achieve accurate performance prediction for each indoor unit of variable-topology VRF systems using limited datasets. Specifically, Kolmogorov-Arnold networks (KANs) are first applied to construct high-fidelity component-level surrogate models for the outdoor unit and indoor units. Subsequently, a physics-informed neural network (PINN) is integrated to predict the system’s evaporator temperature by solving the overall energy balance equation. Finally, the model’s robustness and generalization are validated through interpolation, extrapolation and real-world application tests. Results show a significant enhancement in capturing physical laws in interpolation tests, and the model reduces the root mean square error (RMSE) to 0.24–0.36 kW in extrapolation tests with unseen indoor unit combinations, achieving an accuracy improvement of over 80% compared to the baseline data-driven model. The prediction accuracy (R²) for quasi-steady-state responses in real-world application scenarios exceeds 97%. This physics-guided data-driven modeling framework of VRF systems provides a computationally efficient and physically consistent foundation for the intelligent management of building energy system.