The　dioxin　(DXN)　is　a　key　environmental　indicator　for　municipal　solid　waste　incineration　(MSWI)　process.　However,　the　concentration　of　DXN＇s　emission　prediction　model　always　relies　on　long-term　and　expensive　offline　experimental　analysis.　To　address　this　issue,　some　researchers　employed　deep　neural　networks　(DNN)　to　construct　a　soft　measurement　model,　but　it　has　poor　training　efficiency,　inflexible　model　size,　and　weak　interpretability.　Recently,　the　deep　forest　regression　(DFR)　algorithm　has　achieved　success　in　the　field　of　modeling　domain.　Therefore,　it　is　used　for　DXN　emission　prediction　in　this　paper.　To　encourage　the　diversity　of　features　between　layers,　this　paper　endeavors　to　implement　DFR　for　DXN　emission　concentration　soft　measurement　by　a　new　representation　strategy.　Firstly,　the　dioxin　emissions　and　data　characteristics　of　the　MSWI　process　are　described.　Then,　the　structure　of　DFR　is　analyzed　and　the　problems　of　representation　strategy　are　stated　briefly.　Finally,　we　explore　the　representation　learning　inside　the　DFR　structure　based　on　stacked　generalization.　The　effectiveness　of　the　proposed　method　is　verified　by　DXN　data　of　the　MSWI　process.