Accurate　traffic　flow　forecasting　plays　an　increasingly　important　role　in　traffic　management　and　intelligent　information　service.　Mining　and　analyzing　the　hidden　rules　and　patterns　in　the　historical　data　of　traffic　flow　are　helpful　to　understand　the　rules　of　the　data　and　better　assist　the　prediction.　For　the　long-term　sequence　similarity　measurement,　this　paper　proposes　the　correlation　matrix　sequence　description　method　based　on　wavelet　decomposition,　which　can　better　express　the　sequence　information　and　perform　better　in　the　long-term　prediction　compared　with　Euclidean　distance.　Furthermore,　we　propose　a　similar　search　scheme　based　on　the　nearest　neighbor　and　seasonality.　The　searched　candidates　are　input　into　the　prediction　model　as　the　attention　value,　and　the　output　of　prediction　results　is　assisted　at　each　step.　Compared　with　the　state-of-the-art　methods　on　the　PeMS　dataset,　the　proposed　model　can　effectively　learn　the　long-term　dependence　of　time　series　and　perform　better　in　detail,　showing　advantages　in　multi-step　prediction.