The current distribution network structure is becoming increasingly complex and the paths of disturbance propagation are intertwined. Traditional mechanism modeling based positioning methods have shortcomings in both accuracy and adaptability. Therefore, the paper proposes an improved voltage sag source localization method using variational mode decomposition and hybrid networks. Firstly, in order to extract the key features of voltage sag signals, the paper proposes a VMD voltage sag feature extraction and analysis method that integrates rime optimization, effectively solving the problems of parameter dependence, poor adaptability, and mode aliasing that exist in traditional VMD. Secondly, in response to the issue of segment positioning involving network wide information and high data dimensions, the paper proposes an improved VMD and hybrid network positioning method. Firstly, the improved VMD is used to extract key features from the voltage signals of sparse monitoring points, and then the features are input into a temporal convolutional network lightweight gradient boosting tree model for segment localization. Through numerical verification, the accuracy of localization is above 99.76%.