In　practice,　accurate　and　timely　forecasting　of　short-term　intense　rainfall　is　critical,　but　the　problem　is　extremely　difficult　because　to　its　complicated　spatial-temporal　association.　Although　several　spatial-temporal　series　forecasting　methods　have　been　used　to　rainfall　prediction,　these　models　continue　to　suffer　from　inadequate　modeling　of　data＇s　complicated　intrinsic　connection.　We　provide　a　new　short-term　intense　rainfall　prediction　model　that　use　two　graph　generators　to　model　data　correlations　under　distinct　semantics,　followed　by　a　graph　convolution　module　for　information　integration　to　fully　extract　data　spatial-temporal　information.　Finally,　a　variant　of　recurrent　neural　network　is　employed　to　extract　the　temporal　dependence.　The　experimental　results　on　both　datasets　show　that　the　model　can　model　the　spatial　and　temporal　dependence　across　the　data　more　effectively　than　the　baseline　model,　and　further　improve　the　model＇s　predictive　performance　for　short-term　intense　rainfall.　©　2022　ACM.