Abstract: A precise positioning method based on improved variational mode decomposition and multi domain feature fusion is proposed to address the low accuracy and recognition efficiency of grounding fault location in DC distribution systems. By adaptively determining the number of modes through multi-scale permutation entropy and establishing a dynamic adjustment mechanism for penalty factors, fault features are extracted from the time domain, frequency domain, and time-frequency domain. Relief-F algorithm is used to screen key features, and a fusion strategy combining shallow weighted fusion and deep neural network is designed. Support vector machine is used to identify fault sections, and accurate positioning is achieved by combining dual and traveling wave ranging and error compensation models. The simulation results show that the average positioning error of this method is controlled within 10 m under different transition resistance conditions, with a total time of 119.83 ms, meeting the real-time requirements.