In　the　process　of　training　for　machine　learning,　unbalanced　samples　are　inevitable.　At　the　same　time,　the　cost　of　misclassification　for　stable　samples　and　unstable　samples　is　different.　Thus,　a　cost-sensitive　stacked　variational　auto-encoder　(SVAE)　based　transient　stability　assessment　method　for　power　system　was　proposed.　In　this　paper,　the　tendency　of　the　model　trained　by　unbalanced　sample　was　corrected　by　changing　the　weight　coefficient　of　model　parameter　adjustment.　On　this　basis,　the　weight　coefficient　of　unstable　samples　was　further　improved.　The　fitting　degree　of　the　model　to　the　unstable　sample　was　improved　effectively,　and　the　false　judgment　of　the　unstable　sample　was　reduced.　The　simulation　results　under　the　IEEE　39-bus　system　show　that　the　tendency　of　discriminant　results　under　unbalanced　samples　can　be　improved　and　the　misjudgment　of　unstable　samples　can　be　reduced　by　cost-sensitive　method.　©　2020　Chin.　Soc.　for　Elec.　Eng.