From　the　two　aspects　of　model　construction　and　characteristic　quantities　extraction,　a　transient　stability　discriminant　model　with　better　noise　immunity　is　proposed.　A　stacked　variational　auto-encoder　is　adopted　to　construct　the　assessment　model.　Besides,　a　L2　regularization　method　is　introduced　in　the　trai-ning　process,　which　enhances　the　generalization　ability　of　the　stability　discriminant　model.　Meanwhile,　the　characteristic　quantities　extraction　time　of　the　proposed　method　is　different　from　the　traditional　method.　By　setting　the　threshold　of　the　maximum　power　angle　difference　of　all　generators,　when　the　system　develops　to　the　threshold,　the　characteristic　quantities　extraction　is　carried　out.　The　simulative　results　based　on　IEEE　39-bus　system　show　that　the　miscalculation　of　the　stability　assessment　model　is　greatly　reduced　with　the　proposed　characteristic　quantities　extraction　method.　Meanwhile　the　reasonable　threshold　will　not　affect　the　start　of　real-time　control　methods,　and　the　noise　immunity　ability　of　the　model　can　be　also　strengthened.　©　2019,　Electric　Power　Automation　Equipment　Press.　All　right　reserved.