Abstract: To address the challenges posed by distributed renewable energy sources connecting to the distribution network, synchrophasor measurement has been introduced to the distribution level. The problem that urgently needs to be addressed is how to effectively utilize this massive amount of unlabeled data to identify disturbances and provide data support for power grid operation and control. This paper proposes an unsupervised feature extraction framework called long-short-term time generative adversarial network (LST-TimeGAN) to tackle this problem. The proposed method uses time-series generative adversarial networks (TimeGAN) and introduces an improved framework based on the least squares decision loss function to extract features that can reflect the degree of abnormality of events and provide a basis for accurate classification. Also, a feature extraction unit based on attention mechanism is proposed to improve the efficiency of spatial feature extraction. Furthermore, a long-short three-window parallel framework is established to acquire sensitivity to disturbance features of different time scales. Finally, disturbance identification is completed using a pre-classification and re-identification classification strategy. Verification in simulations and field data shows that this method can accurately identify disturbances even when there are no or few labels. Moreover, it can identify not only disturbances in the transmission network but also local power quality disturbances.