With　the　continuous　enrichment　of　traffic　Internet-of-Things　data　acquisition　methods,　more　and　more　spatiotemporal　data　on　road　networks　is　collected　in　real　time　by　various　sensors　and　multimedia　devices.　The　data-driven　deep　learning　approach　can　make　full　use　of　real-time　data　from　a　road　network　to　predict　future　traffic　status.　By　mining　the　spatiotemporal　relationships　between　road　units,　the　ability　to　predict　network　evolutionary　behaviors　is　improved,　which　provides　a　new　method　of　traffic　management.　There　are　strong　semantic　relations　between　road　intersections　or　road　sections　in　terms　of　traffic　evolution.　Modeling　the　network　only　from　a　shallow　spatial　topological　perspective　ignores　the　important　intrinsic　association　of　the　dynamic　network.　In　this　paper,　we　propose　a　semantic　associative　neural　network　(SANN)　for　traffic　evolution　analysis　by　modeling　the　propagation　effects　and　similarity　patterns　between　road　units.　Considering　the　inadequacy　of　the　fixed　adjacent　matrix,　graph　convolution　is　used　to　encode　the　semantic　features　of　a　road　network　and　embed　them　in　a　bidirectional　recurrent　neural　network　for　sequence　prediction.　Finally,　the　experiments　are　conducted　based　on　speed　data　sets　to　prove　the　effectiveness　of　the　proposed　method.　The　model　achieved　a　well-predicted　accuracy　of　95.33%　and　84.08%　on　Pems-Bay　and　Los　Angeles　data　sets.