To　achieve　precise　subway　passenger　entry　and　exit　flows　prediction,　a　deep　learning-based　for　subway　passenger　flow　pediction　graph　neural　spatio-temporal　network　(GNSTNet),　was　proposed.　Historical　24-hour　subway　entry　and　exit　flow　data,　weather　and　time　labels　were　taken　into　account　as　inputs,　and　they　were　used　to　predict　the　hourly　passenger　entry　and　exit　flows　for　cvcry　Station　in　the　entire　network　for　the　next　hour.　A　subway　passenger　flow　spatio-temporal　graph　were　constructed　using　the　hourly　Station　entry　and　exit　flows　and　subway　network　adjacency　matrix.　In　the　GNSTNet　model,　a　graph　neural　network　was　used　to　extract　the　spatial　features　at　each　time　step.　The　Fourier　transform　was　used　to　extract　the　potential　periodicity　in　the　time　series*　and　a　convolulional　neural　network　was　used　to　extract　the　temporal　features.　The　research　results　show　that　on　a　dataset　from　the　Beijing　Subway　in　June　2021,　the　graph　neural　network-based　subway　passenger　flow　prediction　model　outperforms　six　benchmark　models*　with　an　average　reduction　of　14.　97%　in　the　mean　absolute　error　and　13.　35%　in　the　root　mean　squared　error　compared　to　the　most　accurate　benchmark　model.　The　graph　neural　networks-based　subway　passenger　flow　prediction　model　effectively　improves　the　accuracy　of　predicting　the　passenger　entry　and　exit　flow　for　the　entire　subway　network　compared　to　traditional　algorithms　by　integrating　graph　neural　network,　Fourier　transform　and　convolulional　neural　network.　©　2025　Chang＇an　University.　All　rights　reserved.