As the main form of central air conditioning system, the VRF heat pump air conditioning system has a market share of more than 50%, and the model-based intelligent control technology is a key way to improve its operational energy efficiency. This study selects representative typical cities in the five climate zones and uses the DesignBuilder software to obtain the building load data of each climate zone as the research basis. In terms of modelling approach, this study combines physical models with data-driven methods. On the one hand, a 4R3C load prediction physical model is constructed based on the theory of building RC thermal network, and the intelligent identification of the optimal RC parameters for heating and cooling loads is realised by the genetic algorithm, and the finally obtained CVRMSE indexes are all better than 15%. On the other hand, the multi-layer perceptron (MLP) data-driven prediction model is established based on the machine learning theory, and its prediction performance is verified through systematic training and testing. The comparative study of the two models provides a solid theoretical foundation for the model predictive control of VRF heat pump air conditioning systems.