More　and　more　attention　has　been　paid　to　electronic　medical　record　(EMR)-based　auxiliary　diagnosis　and　treatment,　in　which　medication　recommendation　is　an　important　research　direction.　The　existing　medication　recommendation　models　mainly　depend　on　the　data　of　patients,　diagnosis　and　medications.　However,　the　insufficient　amount　of　clinical　data　with　temporal　dependencies　becomes　a　major　obstacle.　This　paper　proposes　a　new　knowledge-enhanced　pre-training　model　for　medication　recommendation.　On　the　one　hand,　the　classification　knowledge　in　diagnostic　codes　and　drug　codes　is　encoded　by　Graph　Attention　Network　and　fused　into　the　clinical　data　for　expanding　the　data　content.　On　the　other　hand,　a　large　number　of　single　visit　data　of　EMR　are　used　to　create　the　pre-trained　visit　model　by　a　modified　BERT　for　expanding　the　data　scale.　The　experimental　results　on　EMR　data　from　more　than　2,000　medical　and　health　institutions　in　Hainan,　China　show　that　the　fusion　of　classification　knowledge　and　pre-training　model　can　effectively　improve　the　accuracy　of　medication　recommendation.