Due　to　the　emergence　of　new　network　attack　technologies,　cloud　manufacturing　platforms　may　be　subject　to　various　network　attacks　at　any　time.　Traditional　network　attack　prediction　aims　to　predict　the　upcoming　types　of　network　attacks　with　monitored　data　characteristics,　and　this　prediction　method　cannot　accurately　learn　the　network　attack　traffic　that　the　cloud　manufacturing　platform　may　suffer　in　the　future.　Network　attack　traffic　prediction　means　the　future　network　attack　traffic　prediction　by　using　past　data.　Traditional　machine　learning　approaches　cannot　investigate　complex　nonlinear　features.　Deep　learning　methods　can　investigate　nonlinear　characteristics,　yet　they　suffer　from　the　problem　of　overfitting.　In　addition　to　this,　deep　learning　approaches　suffer　from　gradient　disappearance　and　explosion.　To　solve　them,　we　design　a　network　attack　traffic　method　called　S-Informer,　which　integrates　the　filter　of　Savitzky-Golay,　the　self-attention　of　ProbSparse　in　a　generative　decoder　and　an　encoder　for　eliminating　noise,　reducing　the　network　scale　and　improving　the　speed　of　prediction.　Real-life　dataset-based　experimental　results　show　that　S-Informer　achieves　higher　prediction　accuracy　than　several　commonly-used　algorithms.