Deep learning technology is identified as a valid tool for transient stability assessment (TSA). Moreover

the superior performance of the TSA model depends on generously labeled samples. However

the power grid is dynamic

and some topologies or operation conditions change substantially. The traditional method generates a significant quantity of samples for each specific topology. Nonetheless

generating these labeled samples and establishing TSA models is very time-consuming. This paper proposes a high-quality sample generation framework based on data-driven methods to build a high-quality offline samples database for TSA model training and updating. Firstly

the representative topologies provided by the system operator are clustered into four different categories by density-based spatial clustering of applications with noise (DBSCAN). Thus the corresponding samples are collected. Then

when a new topology is encountered in the online application

scenario matching is used to match the most similar topology category. After that

instance-based transfer learning is implemented from a database of the best-matched topology category. Finally

a deep convolutional generative adversarial network (DCGAN) is constructed to mitigate the class imbalance problem. That is

unstable scenarios occur far more rarely than stable scenarios. Consequently

a high-quality and balanced TSA model training and updating database is constructed. The comprehensive test results on the Central China Power Grid illustrate that the proposed framework can generate high-quality and balanced TSA samples. Furthermore

the sample generation time is dramatically shortened. In addition

the metrics of accuracy

reliability and adaptability of the TSA model are significantly enhanced.