Transient　stability　assessment　(TSA)　is　usually　considered　as　a　challenging　problem　in　power　systems.　During　the　measurement　and　transmission　of　real-time　data　in　power　grids,　different　degrees　of　noise　may　occur.　The　real　noise　is　complex　and　hard　to　predict,　so　the　current　methods　cannot　output　satisfied　TSA　results.　Therefore,　an　adaptive　denoising　combined　model　(ADCM)　is　proposed　in　this　paper.　The　ADCM　can　adaptively　output　the　results　of　TSA　in　complex　noise　environment.　Firstly,　the　stacked　denoising　auto-encoder　based　model　trained　by　data　with　expected　noise　is　proposed,　which　is　called　targeted　denoising　model　(TDM).　The　TDM　can　output　satisfied　results　for　specific　noisy　data.　Then,　the　ADCM　composed　of　multiple　TDMs　and　a　noise　recognition　model　is　constructed.　The　recognition　model　is　a　regression　model　based　on　deep　neural　networks.　When　actual　data　are　input,　the　recognition　model　can　output　weight　values　for　TDMs.　And　when　the　actual　noise　is　similar　to　the　noise　trained　for　the　TDM,　the　weight　obtained　is　great　for　this　model.　The　final　results　are　obtained　by　assigning　the　weights　to　the　results　of　each　TDM.　The　effectiveness　of　this　method　is　verified　by　simulation　results　in　IEEE　39-bus　system　and　realistic　system.