Traffic　speed　prediction　in　the　urban　road　network　is　an　important　and　promising　task　of　intelligent　transportation　systems.　Precise　traffic　speed　prediction　can　mitigate　traffic　congestion　and　improve　road　network　utilization.　This　task　is　challenging　because　of　the　complexity　of　the　spatiotemporal　dependency　of　traffic　data　among　road　network.　Existing　approaches　mainly　focus　on　the　whole　road　network　and　it　may　capture　much　redundant　information　and　lead　to　high　computational　cost.　In　this　paper,　we　propose　a　Key　Path-Based　Deep　Learning　Approach:　Path-Based　CNN-1D　+　GRU　+　CNN-2D　(P-CGC),　a　novel　deep　learning　model　for　traffic　speed　prediction.　Specifically,　we　use　EST-matching　algorithm　to　match　the　float　car　data　into　the　road　network.　Then　we　select　several　key　paths　and　build　the　model　for　the　loop　detectors　which　are　in　the　same　key　path.　We　introduce　CNN-1D,　GRU　to　extract　the　temporal　dependency　of　the　data,　where　CNN-1D　is　used　to　fuse　the　contextual　information,　and　GRU　is　used　to　capture　the　features　of　the　temporal　dimension.　Then　we　concatenate　the　output　of　all　CNN-1D+GRU　models　and　use　CNN-2D　to　capture　the　spatial　dependency　of　the　data.　Finally,　a　fully　connected　neural　network　is　used　to　transform　features　into　the　prediction.　We　conduct　extensive　experiments　on　Zhangzhou　real-world　datasets,　and　the　proposed　approach　achieves　a　good　result.　©　Published　under　licence　by　IOP　Publishing　Ltd.