Fast　and　reliable　transient　stability　assessment　(TSA)　is　significant　for　safe　and　stable　power　system　operation.　Deep　learning　provides　a　new　tool　for　TSA.　However,　it　is　difficult　to　apply　the　TSA　models　based　on　deep　learning　practically　because　of　their　inexplicability.　Therefore,　an　interpretable　time-adaptive　model　based　on　a　dual-stage　attention　mechanism　and　gated　recurrent　unit　(GRU)　is　proposed　for　TSA.　A　feature　attention　block　and　a　time　attention　block　are　included　in　the　dual-stage　mechanism　to　explain　the　TSA　rules　learned　by　the　proposed　TSA　model.　Meanwhile,　interpretability　is　utilized　to　guide　the　optimization　of　the　TSA　model.　Firstly,　the　measurements　are　input　into　the　feature　attention　block　to　calculate　feature　attention　factors.　Then,　the　measurements　weighted　by　the　feature　attention　factors　are　input　into　a　GRU　block　for　further　abstracting.　The　abstracted　features　are　input　into　the　time　attention　block　to　obtain　time　attention　factors.　Finally,　the　abstracted　features　weighted　by　the　time　attention　factors　are　sent　into　fully　connected　layers　for　TSA.　To　achieve　time-adaptive　TSA,　multiple　channels　are　constructed　to　process　the　features　at　different　decision　moments.　The　performance　of　the　proposed　model　is　verified　in　the　IEEE-39　bus　system　and　a　realistic　regional　system.