To　predict　the　oxygen　content　in　flue　gas　quickly　and　accurately　during　the　municipal　solid　waste　incineration　(MSWI),　the　dynamic　pruning　strategy　based　on　mutual　information　(MI)　is　used　to　build　a　learner　model　of　deep　stochastic　configuration　network　(DSCN)　in　this　paper.　This　learning　model　consists　of　two　parts.　One　is　to　select　the　characteristic　variables　of　the　oxygen　content　in　flue　gas　through　the　improved　sailfish　optimizer　(SFO)　with　t-distribution.　The　other　is　to　use　MI　to　prune　hidden　nodes　dynamically　in　a　layer-by-layer　manner　during　the　training　of　DSCN,　and　then　obtain　a　prediction　model　of　the　oxygen　content　in　flue　gas.　The　experimental　results　show　that　the　improved　SFO　can　select　the　optimal　characteristic　variables,　MI　can　effectively　reduce　the　complexity　of　the　DSCN　after　pruning　and　achieve　the　rapid　and　accurate　prediction　of　oxygen　content　in　flue　gas,　which　lays　a　foundation　for　timely　optimization　and　adjustment　of　incineration　conditions.