In　order　to　achieve　the　purpose　of　reducing　the　loss　of　the　offshore　wind　power　transmission　project,　the　paper　deduces　the　transmission　efficiency　function　of　the　submarine　cable　in　the　offshore　wind　power　transmission　project　in　detail,　analyzes　the　influencing　factors　of　the　transmission　efficiency　of　the　submarine　cable　and　the　loss　reduction　principle　of　the　optimization　for　the　on-load　transformer　tap　changer,　and　establishes　the　daily　loss　optimization　model　of　the　submarine　cable　transmission　project.　Then,　given　the　time　series　auto-correlation　analysis　of　random　variables,　a　predictive　data-driven　approach　to　modeling　a　high-dimensional　convex　hull　uncertain　set　is　applied　in　this　paper　to　consider　the　fluctuations　in　the　wind　farm　output　and　the　grid　network　voltage　in　the　loss　reduction　optimization,　which　gets　extreme　scenarios　and　reduces　the　decision　conservatism　while　improving　the　computational　efficiency.　In　addition,　the　Q-learning　algorithm　is　used　for　training　based　on　extreme　scenarios　obtained　and　combined　with　the　multi-agent　system　optimization　method　to　solve　the　optimization　model　by　the　tap　position　optimization　and　adjustment　time　optimization　alternately.　The　agent　learns　from　the　interaction　between　the　action　strategy　and　the　wind　farm　state,　so　as　to　form　the　best　tap　changer　setting　and　the　best　adjustment　time　strategy　under　various　operating　states　of　the　offshore　wind　farm.　Finally,　the　rationality　of　the　proposed　model　and　the　validity　of　the　method　are　verified　through　the　simulation　analysis　of　a　220　kV　AC　submarine　cable　transmission　project　in　an　offshore　wind　farm.　©　2024　Chin.Soc.for　Elec.Eng.