The three-level T-type shunt active power filter (3LT²SAPF) has been widely used to improve the power quality of power grid under nonlinear load, and its harmonic suppression and reactive power compensation performance and the corresponding reliability have attracted wide attention. Meanwhile, model predictive control (MPC) can achieve multi-objective optimization. However, once the parameter mismatch occurs due to reasons such as changes in temperature, the output performance of an MPC system will be seriously deteriorated. This may even may endanger the system stability. Additionally, when the circuit relation is unknown, the establishment of a predictive model is not available, and the MPC controller will not be applicable. To solve these problems and enhance the robustness of the 3LT²SAPF system, a linear fitting sequential model-free predictive control method was proposed. First, the principle of linear fitting method for current prediction was analyzed. Second, a current error matrix for each switching vector was established, which can be used for harmonic current tracking. Finally, the cost functions for neutral-point (NP) voltage and DC-bus voltage were designed. The results of numerous experiments under different scenarios verified the correctness and effectiveness of the proposed method.