Dioxin　(DXN)　emission　concentration　is　an　important　environmental　indicator　in　the　municipal　solid　waste　incineration　(MSWI)　process.　The　prediction　model　of　DXN　emission　can　be　used　for　pollution　control　to　realize　actual　requirements　of　operation　optimization.　Therefore,　a　DXN　emission　concentration　prediction　model　based　on　improved　deep　forest　regression　(ImDFR)　is　proposed　in　this　study.　A　feature　reduction　layer　based　on　out-of-bagging　error　is　first　introduced　into　the　ImDFR　to　eliminate　redundant　variables　and　feed　all　confidence　information　on　DXN　emission　into　the　feature　enhancement　layer　of　the　MSWI　process.　A　deep　ensemble　stacking　model　is　subsequently　built　to　depict　deep　features　and　increase　diversity　and　accuracy　using　random　forests,　completely　random　forests,　GBDT,　and　XGBoost　as　subforests.　Finally,　the　predicted　value　of　the　DXN　prediction　model　is　determined　in　the　decision　layer.　The　DXN　emission　prediction　model　is　verified　using　actual　historical　data　of　two　incinerators　operated　with　a　daily　processing　capacity　of　800　tons.　The　experimental　results　showed　that　the　proposed　prediction　model　presents　higher　accuracy　and　better　generalization　ability　than　state-of-the-art　models.