Abstract: To address the reduced reliability of fault location caused by complex branches, transient noise, fault resistance and variable operating conditions in distribution networks, a fast fault location method based on improved transient feature extraction is proposed. The method combines parameter-adaptive variational mode decomposition with wavelet packet energy entropy to extract multi-scale transient current features, and uses a lightweight extreme learning machine for faulty-section identification and distance estimation. The proposed framework balances feature interpretability and online computational efficiency, providing an engineering-oriented implementation scheme for coordinated fault location between distribution automation terminals and master stations.