Recently,　data-driven　methods　have　been　widely　assessed　by　researchers　in　the　field　of　power　system　transient　stability　assessment　(TSA).　The　differences　in　prediction　difficulty　among　the　samples　are　ignored　by　most　previous　studies.　To　address　this　problem,　anchor　loss　(AL)　is　introduced,　which　can　dynamically　reshape　loss　values　based　on　the　prediction　difficulty　of　samples.　Thereby,　easy　samples　are　suppressed　by　reducing　their　loss　values　to　avoid　being　paid　too　much　attention　when　they　are　misclassified.　Meanwhile,　hard　samples　are　emphasized　by　increasing　their　loss　values,　in　order　to　be　predicted　correctly　as　much　as　possible.　On　basis　of　the　AL,　historical　information　in　the　model　training　process　is　considered.　A　novel　loss　function　named　historical　information　anchor　loss　(HIAL)　is　designed.　The　loss　values　can　be　adaptively　rescaled　according　to　the　previous　prediction　results　as　well　as　the　prediction　difficulty　of　samples.　Finally,　the　HIAL　is　combined　with　the　deep　brief　network　(DBN)　and　applied　in　the　IEEE　39-bus　system,　and　a　realistic　system　is　produced　to　verify　its　effectiveness.　By　incorporating　prediction　difficulty　and　historical　training　information,　the　accuracy　(with　a　reduction　in　misjudgment　rate　exceeding　30%)　and　convergence　speed　of　the　TSA　model　can　be　significantly　improved.