Reliable　transient　stability　assessment　(TSA)　can　support　the　grid　to　adopt　reasonable　emergency　control　measures　timely.　Therefore,　a　time-adaptive　ensemble　TSA　model　is　proposed　to　evaluate　the　stability　accurately　and　rapidly.　Firstly,　a　variable-length　training　set　is　generated　according　to　swing　angles,　which　can　drop　useless　information　and　keep　enough　valid　information.　Secondly,　a　gated　recurrent　unit　(GRU)　network　is　introduced　to　learn　variable-length　temporal　data.　Thirdly,　a　cost-sensitive　method　is　combined　with　the　GRU　network　to　adjust　its　evaluation　tendency.　Then,　two　cost-sensitive-based　GRU　networks　with　opposite　evaluation　tendencies　and　one　GRU　network　without　evaluation　bias　are　integrated.　If　the　final　evaluation　time　is　not　reached　or　the　swing　angle　threshold　is　not　triggered,　the　creditability　of　current　predictions　is　judged　by　the　consistency　of　the　results　predicted　by　the　two　GRU-based　models　with　opposite　preferences.　The　credible　results　can　be　output　directly,　and　the　incredible　results　are　shelved.　If　the　end　mechanism　is　triggered,　the　stability　is　determined　by　the　model　without　preference.　High　accuracy　can　be　achieved　by　lengthening　the　variable-length　data　and　untimely　emergency　control　measurements　can　be　avoided.　The　performance　of　the　proposed　strategy　is　tested　in　the　IEEE-39　bus　system　and　a　realistic　regional　system.