Accurate　and　fast　transient　stability　assessment　is　helpful　for　grid　operators　to　take　effective　emergency　control　actions　after　faults.　The　gradual　deployment　of　the　wide-area　measurement　system　provides　a　basis　for　introducing　machine　learning　into　transient　stability　assessment　(TSA).　However,　the　application　of　the　machine　learning　model　is　restricted　by　the　imbalance　of　samples　in　actual　systems.　In　this　paper,　a　time-adaptive　assessment　model　with　imbalance　correction　based　on　the　ratio　of　loss　function　values　is　built　to　realize　accurate　and　fast　TSA.　First,　a　long　short-term　memory　(LSTM)-based　model　whose　optimization　goal　is　to　reduce　the　prediction　loss　at　a　fixed　time　step　is　established.　Several　LSTM-based　models　with　different　decision　time　values　are　integrated　as　a　multiple　LSTM　(MLSTM)　TSA　model.　It　is　found　that　the　effect　of　imbalanced　samples　on　model　parameters　can　be　quantified　by　the　loss　function　values.　Then,　the　ratio　of　loss　function　values　of　two　classes　is　obtained　by　pre-training,　by　which　the　imbalance　degree　of　data　can　be　quantified.　The　correction　coefficient　is　determined　and　used　to　retrain　LSTMs　to　solve　the　evaluation　tendency　problem.　Simulation　results　in　an　IEEE　39-bus　system　and　an　actual　power　system　show　the　excellent　performance　of　the　proposed　imbalance　correction　scheme　and　evaluation　scheme.　Compared　with　traditional　methods,　imbalance　correction　can　be　achieved　with　better　results.