Transient stability assessment (TSA) plays a critical role in ensuring
the reliable operation of power systems. However, existing approaches
for TSA often encounter challenges such as data imbalances, limited
sample sizes, and the need for adaptability in the face of system
changes, necessitating the exploration of more advanced techniques. This
paper proposes a novel deep transfer learning (DTL) framework to address
these limitations that incorporates CNN-LSTM and stacked denoising
auto-encoder (SDAE) techniques, aiming to significantly improve the
speed and accuracy of power system TSA, especially in online
applications and adaptability to system changes. First, the utilization
of SDAE enables effective feature extraction, while the implementation
of class weight balancing and cross-entropy loss function techniques
effectively addresses data imbalances. Second, a CNN-LSTM classifier is
constructed using transfer progressive learning. This approach allows
for the effective analysis of spatial and temporal dynamic measurements
by leveraging unsupervised pre-training (auto-encoder) and additional
CNN-LSTM layers. Third, we propose the DTL, which leverages knowledge
transfer from the CNN-LSTM model and incorporates fine-tuning
techniques. This innovative approach ensures adaptability under in four
scenarios, which is a prevalent challenge in power systems for
continuous prediction. As compared with other techniques, the results
demonstrate that our proposed approach achieves TSA accuracy of up to
99.68% on the IEEE 39-bus system and 99.80% on the South Carolina
500-bus system. Furthermore, to compare the performance of continuous
prediction with other methods, our proposed method exhibits a
significant improvement of 2% even with a limited sample size.