Abstract: The modular multilevel matrix converter(M3C) has promising application prospects in renewable energy transmission scenarios, such as low-frequency transmission and offshore wind power grid connection. However, due to the complex topology and strong internal coupling, the M3C may suffer from stability problems such as low-frequency oscillations and harmonic resonance under complex operating conditions. Impedance-based analysis can characterize the interaction between the M3C and the power grid from the frequency-domain perspective, thereby providing an effective tool for grid-connected stability assessment. Nevertheless, conventional analytical impedance modeling methods have difficulty in simultaneously ensuring modeling accuracy and adaptability to multiple operating conditions. To overcome this limitation, a neural-network-based impedance modeling method is proposed for the M3C enhanced by feature-distance-based confidence evaluation. In the proposed method, operation condition-impedance data obtained by frequency scanning at multiple operating points are used as training samples, and a data-driven impedance model of the M3C is established through supervised learning. Furthermore, the distance in the deep feature space is employed to evaluate and calibrate the reliability of the prediction results, thus improving the robustness of impedance modeling under different operating conditions. In addition, an impedance dataset covering multiple operating points is constructed to enhance the adaptability of the model to operating-condition variations. Finally, the modeling accuracy of the proposed method is verified using experimental data. The results show that the proposed method can effectively characterize the impedance characteristics of the M3C under multiple operating conditions.