In order to improve the accuracy of power system transient stability assessment

overcome the problems of difficulty in characteristics capture and local convergence of the learning model due to the imbalance of training data

this paper proposes a power system transient stability assessment method based on unconventional sample mining and alternate learning. Firstly

in the face of massive high-dimensional state data

an unconventional sample extraction model based on mutual information entropy and K-means clustering is proposed

which achieves the screening of unconventional samples through key feature distillation and unconventional state seeking. Secondly

based on the screening results

a single-agent transient stability judgement model with alternate learning between full sample and unconventional sample is established

which can improve the neural network's ability to capture the unconventional data features. At the same time

the parallel stability judgement method is adopted to achieve the effective use of various types of single-intelligence

and the overall performance of the stability judgement is improved. Finally

the proposed model is simulated based on the PSD-BPA time-domain simulation platform and the IEEE-39 node system. Compared with the traditional single-agent model

the improved method achieves a 0.53% increase in average accuracy and a 2.34% reduction in average misjudgment rate across multiple test sample groups for transient stability assessment. The simulation results validate the feasibility and effectiveness of the improved transient stability assessment method.