In modern smart grid operation, there is a great imbalance between transient stable samples and unstable samples, and the cost of misjudgment of unstable samples is different from that of misjudgment of stable samples. Current transient stability assessment methods based on machine learning are mostly based on shallow models, which pay insufficient attention to misjudged transient unstable samples and the assessment accuracy needs to be improved. Based on this problem, a rough intensive power system transient stability assessment method is proposed. The neighborhood rough set was used to find multiple optimal feature subsets at different granularity levels to recharacterize the original feature, and the nonlinear mapping between feature quantity and transient stable state was strengthened by machine learning. Weight classification is introduced to improve the attention of transient unstable samples in the classification process. Experimental results on ieee 10-machine 39-node system show that the proposed method can not only improve the evaluation accuracy, but also effectively reduce the misjudgment of transient unstable samples, and has a good performance on unbalanced samples, with certain robustness.