Transient　stability　assessment　(TSA)　of　power　systems　based　on　deep　learning　faces　practical　implementation　challenges　due　to　the　unpredictability　of　evaluation　results　and　uncontrollability　of　decision-making　processes.　While　attention　mechanisms　show　promising　potential　in　addressing　unpredictable　and　uncontrollable　problems,　current　research　primarily　focuses　on　the　former　issue.　To　bridge　this　gap,　this　study　proposes　a　method　that　utilizes　system　dominant　patterns　to　guide　models　in　assigning　more　rational　feature　attention　weights,　thereby　controllably　enhancing　model　generalization　capability.　First,　the　improved　MeanShift　algorithm　is　used　to　cluster　the　generators　of　each　training　set　sample　and　the　critical　cluster　is　labeled　to　capture　the　dominant　pattern.　Then,　an　objective　function　fused　with　dominant　pattern　information　is　constructed　to　optimize　the　distribution　of　attention　weights.　Finally,　the　new　objective　function　is　applied　for　training　and　updating　of　the　model.　The　examples　of　IEEE　39-bus　system　and　East　China　power　grid　show　that　the　model　constructed　using　the　proposed　method　has　stronger　generalization　ability　and　better　noise　resistance,　and　the　model’s　unpredictability　and　uncontrollability　can　be　improved.　©2025　Chin.Soc.for　Elec.Eng.