With　the　global　outbreak　of　COVID-19,　an　increasing　number　of　countries　have　made　imported　epidemic　control　a　priority,　imposing　restriction　measures　to　prevent　the　spread　of　the　virus　caused　by　imported　cases.　To　control　the　imported　epidemic,　it　is　necessary　to　accurately　predict　the　number　of　imported　cases　from　different　source　countries.　This　paper　proposes　a　novel　time　series　prediction　approach　called　PNICA　(Prediction　on　Number　of　Imported　CAses)　that　uses　deep　learning　to　predict　the　number　of　COVID-19　imported　cases.　On　the　one　hand,　the　proposed　PNICA　approach　adopts　a　multi-modal　learning　strategy　to　fuse　three　sources　of　data:　flight　data,　the　epidemic　data,　and　the　data　of　historical　imported　cases.　On　the　other　hand,　the　proposed　PNICA　approach　extends　the　traditional　transformer　model　with　cross-modal　attention　to　learn　the　interactions　between　different　data　modalities　to　improve　prediction　accuracy.　We　use　China　as　the　target　country　and　collect　the　number　of　imported　cases　from　four　source　countries-Japan,　USA,　Russia,　and　the　UK-as　well　as　the　epidemic　data　and　flight　data　from　May　to　November　2020.　Experiments　on　the　collected　data　demonstrate　that　the　proposed　PNICA　approach　outperforms　the　baseline　methods　in　predicting　the　number　of　imported　cases.　The　ablation　study　shows　that　both　the　multi-modal　learning　strategy　and　cross-modal　attention　can　significantly　improve　prediction　performance.