Network　traffic　has　apparent　characteristic　of　burst.　The　time　series　of　it　presents　nonlinear.　It　is　difficult　for　traditional　linear　method　to　predict　it　accurately.　To　solve　the　problem,　this　paper　proposes　to　decompose　the　original　traffic　to　an　approximation　sequence　and　several　detail　sequences　with　the　method　of　wavelet　transformation.　On　this　basis,　the　change　trend　of　traffic　is　learned　by　LSTM　network　and　the　burst　information　is　extracted　at　multiscale　to　complete　the　prediction　of　future　traffic.　The　experimental　results　show　that　for　the　prediction　error,　the　model　constructed　with　LSTM　network　is　superior　to　the　models　constructed　with　LSSVM,　BP　neural　network　and　Elman　neural　network.　In　addition,　the　model　proposed　in　this　paper　performs　better　than　the　ordinary　LSTM　network　model　for　predicting　the　burst　of　traffic.　©　2018　IEEE.