Accurately identifying voltage sag sources is crucial for determining responsibility and control entities in power quality disturbances. However, in real power grids, voltage sag events often have limited and imbalanced samples, which limits the performance of supervised learning classification models. To solve this, a voltage sag responsibility allocation method for distribution networks based on the WGAN - GP algorithm is proposed. First, train a WGAN - GP model for each sag source using a small amount of real sag waveform data to learn the data distribution. Then, use the trained generator to expand the dataset by creating numerous high - quality sag samples. Finally, train a CNN classifier on this enhanced dataset to accurately identify sag sources and allocate responsibility. Case studies on a distribution network model in the Simulink platform show that the WGAN - GP algorithm effectively handles small sample problems, offering a new technical approach for voltage sag responsibility allocation.