The　real-time　transient　stability　assessment　(TSA)　is　critical　for　emergency　control　of　power　systems.　Accurate　and　fast　TSA　can　provide　an　important　basis　for　post-fault　control　of　power　systems.　At　present,　the　accuracy　of　the　evaluation　model　based　on　machine　learning　is　very　high,　but　there　are　still　some　misjudgements　in　the　results.　In　order　to　build　a　high-accuracy　evaluation　model,　a　novel　frame　based　on　the　cost-sensitive　method　is　proposed.　Firstly,　a　deep　belief　network　(DBN)　is　applied　to　build　a　TSA　frame.　The　DBN　is　effectively　trained　by　means　of　pre-training　and　fine　tuning.　Secondly,　two　models　with　the　opposite　preference　are　constructed　based　on　the　cost-sensitive　method.　Based　on　the　output　results　of　the　two　models,　the　deterministic　or　uncertain　evaluation　results　are　obtained.　The　samples　that　may　be　misclassified　are　divided　into　uncertain　evaluation　results.　Thus,　the　accuracy　of　the　deterministic　evaluation　results　can　be　improved　greatly.　Through　this　design,　the　practicability　of　the　evaluation　model　based　on　machine　learning　is　greatly　improved.　The　effectiveness　of　the　proposed　scheme　is　verified　by　the　simulation　results　in　the　IEEE-39　bus　system　and　a　realistic　regional　system.