Deep　learning　methods　are　widely　adopted　in　power　system　transient　stability　assessment　(TSA).　However,　the　interpretability　of　the　assessment　results　and　the　controllability　of　the　assessment　process　hinder　the　further　application　of　deep　learning　methods　in　practice.　In　this　paper,　an　instability　pattern-guided　model　updating　method　is　proposed　to　optimize　the　TSA　model.　Firstly,　a　TSA　model　based　on　Transformer　encoder　is　proposed　to　explain　and　analyze　the　model's　prediction　through　attention　distribution.　Secondly,　an　attention-guiding　loss　is　employed　to　revise　the　assessment　rules　for　specified　instability　patterns.　The　samples　with　specified　instability　patterns　can　be　classified　more　accurately.　Thirdly,　an　attention-keeping　loss　is　employed　to　maintain　the　assessment　rules　for　other　samples　and　mitigate　overfitting　in　the　update.　In　addition,　a　representative　dataset　is　introduced　to　reduce　the　update　cost.　The　samples　in　the　representative　dataset　are　extracted　from　an　original　training　set　based　on　the　attention　distribution.　The　effectiveness　of　the　proposed　method　is　verified　in　the　IEEE　39-bus　system　and　the　East　China　Power　Grid　system.　IEEE