Short-term　traffic　states　forecasting　of　road　networks　based　on　real-time　data　is　an　important　component　of　intelligent　transportation　systems,　especially　advanced　traffic　management　systems　and　traveller　information　systems.　By　considering　the　influence　of　both　space　and　time　dimensions,　we　proposed　a　novel　GATs-GAN　framework　for　the　forecasting　of　traffic　states.　First,　to　capture　spatial　traffic　relationships,　the　traffic　topological　graph　network　is　set　up　based　on　the　connection　of　traffic　sections.　Then,　the　first-order　neighbours　and　high-order　neighbours　of　traffic　networks　can　be　structured.　Graph　attention　networks　(GATs)　are　used　to　obtain　the　hidden　features　of　input　traffic　data　by　training　the　attention　between　nodes　in　high-order　neighbours.　Based　on　two　traffic　networks　in　California　and　Seattle　in　the　United　States,　we　find　that　the　GATs-GAN　with　high-order　neighbours　can　satisfactorily　estimate　the　traffic　data　and　performs　better　than　the　baseline　methods　and　comparative　experiments.