Accurate　traffic　flow　prediction　is　becoming　increasingly　important　for　transportation　planning,　control,　management,　and　information　services　of　successful.　Numerous　existing　models　focus　on　short-term　traffic　forecasts,　but　effective　long-term　forecasting　of　traffic　flows　have　become　a　challenging　issue　in　recent　years.　To　solve　this　problem,　this　paper　proposes　a　deep　learning　architecture　which　consisting　of　two　parts:　the　long　short-term　memory　encoder-decoder　structure　at　the　bottom　and　the　calibration　layer　at　the　top.　In　the　encoder-decoder　model,　we　propose　an　hard　attention　mechanism　based　on　learning　similar　patterns　to　enhance　neuronal　memory　and　reduce　the　accumulation　of　error　propagation.　To　correct　some　of　the　missing　details,　we　design　a　control　gate　in　the　calibration　layer　to　learn　the　predicted　data　in　groups　according　to　different　forms.　The　proposed　method　is　evaluated　on　real-world　datasets　and　compared　with　other　state-of-the-art　methods.　It　is　verified　that　our　model　can　accurately　learn　local　feature　and　long-term　dependence,　and　has　better　accuracy　and　stability　in　long-term　sequence　prediction.