Variable refrigerant flow (VRF) systems are widely utilized in public buildings. To enhance their energy efficiency, existing studies predominantly focus on equipment-side optimization, with fewer strategies targeting user-side optimization. This study proposes an energy-saving strategy for indoor set temperature based on data-driven predictive control, considering the dynamic balance among energy consumption, electricity costs, and thermal comfort. A LightGBM model is trained to predict the system’s energy consumption and average indoor temperature over a specific time horizon. The optimization problem, aiming to minimize energy consumption and indoor temperature rise, is solved using the differential evolution algorithm to determine the optimal set temperature. A VRF simulation platform is developed using Python and EnergyPlus to validate and analyse the energy-saving strategies under different temperature rise penalty coefficients. The results show that the model accurately predicts energy consumption and average indoor temperature. Compared with maintaining a fixed set temperature of 25 ℃, the proposed strategy achieves up to 7.49% energy savings and 8.22% electricity cost savings.