Abstract: In medium- and low-voltage DC distribution systems, the direct current transformer (DCT) often adopts model predictive control (MPC) to improve the dynamic response characteristics of the system. However, its strong parameter dependence and unbalanced transmission power are the key factors limiting the development of MPC. A model-free predictive control (MFPC) method is proposed for this purpose, which has the advantages of parameter insensitivity and transmission power self-balancing. Firstly, by establishing ultra-local model for a single dual active bridge (DAB), lumped disturbance is analyzed and parameters for passive components and unmodeled parts are calculated in real-time, thereby enhancing the robustness of the control system. Subsequently, the lumped disturbance and input voltage balancing are integrated into the discrete model of the output voltage, enabling the DCT to improve its output voltage accuracy and power balancing capabilities under parameter mismatch conditions without incurring additional computational costs. Finally, a 120 V/600 W experimental prototype was built to verify the effectiveness and superiority of the proposed control method.