Modern　blast　furnace　ironmaking　technology　mainly　uses　the　thermal　state　of　the　furnace　cylinder　to　reflect　the　furnace　temperature　conditions.　However,　due　to　the　complexity　of　the　blast　furnace　smelting　process,　it　is　very　difficult　to　modelling　and　control　this　process　effectively.　Therefore,　it　is　important　to　carry　out　research　on　blast　furnace　temperature　prediction　modelling　in　order　to　realize　early　warning　of　furnace　health　condition　in　production　systems,　where　nowadays　more　and　more　attentions　are　paid　to　technologies　of　machine　learning　and　deep　learning.　Taking　the　actual　application　scenario　of　a　large　iron　and　steel　production　enterprise　as　a　case　study,　this　paper　focuses　on　the　lack　of　robustness　when　training　machine　learning　models,　due　to　the　noise　in　the　original　collected　business　data.　The　experimental　results　show　that　the　application　of　feature　engineering,　including　feature　construction,　key　feature　analysis,　and　feature　ranking　within　different　data　analysis　stages,　is　able　to　improve　the　quality　of　the　collected　raw　business　data,　consequently　it　is　helpful　in　solving　practical　engineering　problems　from　machine　learning　perspective.　©　2023　Copyright　held　by　the　owner/author(s).　Publication　rights　licensed　to　ACM.