In　order　to　solve　the　problem　of　evaluation　tendency　caused　by　imbalanced　samples,　the　influence　of　imbalanced　samples　on　evaluation　model　is　analyzed　from　the　loss　function　of　deep　learning　model,　and　it　is　found　that　the　loss　function　value　in　the　training　process　can　reflect　the　imbalance　degree　of　the　samples,　thus　a　cost-sensitive　correction　method　based　on　sample　posterior　distribution　information　is　proposed.　The　posterior　distribution　information　of　the　samples　is　obtained　by　the　preliminary　training,　and　the　correction　coefficient　is　obtained　by　introducing　the　mean　value　ratio　of　loss　function　of　stable　samples　to　unstable　samples.　The　correction　coefficient　is　used　to　correct　the　loss　function　of　the　model　by　the　cost　sensitive　method,　and　the　model　is　trained　again　so　as　to　correct　the　evaluation　tendency.　Compared　with　the　traditional　methods,　the　proposed　method　quantifies　the　imbalance　degree　of　the　model　from　the　training　mechanism,　and　the　correction　coefficient　comprehensively　considers　the　influence　of　the　imbalance　of　sample　quantity　and　spatial　distribution　on　model　parameters,　which　realizes　better　correction　effect.　The　effectiveness　of　the　proposed　method　is　verified　by　the　simulative　results　of　IEEE　39-bus　system　and　a　regional　system　in　East　China.　©　2022,　Electric　Power　Automation　Equipment　Press.　All　right　reserved.