The　accurate　short-term　traffic　flow　prediction　can　provide　timely　and　accurate　traffic　condition　information　which　can　help　one　to　make　travel　decision　and　mitigate　the　traffic　jam.　Deep　learning　(DL)　provides　a　new　paradigm　for　the　analysis　of　big　data　generated　by　the　urban　daily　traffic.　In　this　paper,　we　propose　a　novel　end-to-end　deep　learning　architecture　which　consists　of　two　modules.　We　combine　convolution　and　LSTM　to　form　a　Conv-LSTM　module　which　can　extract　the　spatial-temporal　information　of　the　traffic　flow　information.　Furthermore,　a　Bi-directional　LSTM　module　is　also　adopted　to　analyze　historical　traffic　flow　data　of　the　prediction　point　to　get　the　traffic　flow　periodicity　feature.　The　experimental　results　on　the　real　dataset　show　that　the　proposed　approach　can　achieve　a　better　prediction　accuracy　compared　with　the　existing　approaches.　©　2017　IEEE.