Dioxin　(DXN)　emission　concentration　is　a　key　environmental　index　in　the　process　of　municipal　solid　waste　incineration　(MSWI).　Realizing　an　intelligent　optimization　operation　depends　on　building　a　soft　sensor　model　of　DXN　emission.　However,　data　on　the　MSWI　process　have　missing　and　abnormal　values,　which　lead　to　incomplete　data　used　for　modeling.　In　addition,　soft　sensor　modeling　of　difficult-to-measure　parameters　with　large　lag　characteristics　has　multi-scale　problems.　Therefore,　DXN　emission　soft　sensor　modeling　has　to　tackle　problems　such　as　missing　data,　data　scale　mismatch,　and　small　samples.　This　study　proposes　a　multi-scale　missing　data　modeling　using　improved　generative　adversarial　network　(GAN)　and　deep　forest　regression　(DFR)　to　address　the　abovementioned　problems.　First,　the　short　time-scale　samples　and　features　are　divided　according　to　the　missing　data.　Second,　the　missing　data　are　filled　by　GAN,　in　which　the　MSE　constraint　term　ensures　the　filling　effect　and　the　overfitting　problem　is　alleviated　by　using　the　model　complexity　penalty　term.　Third,　a　certain　range　of　the　short　time-scale　input　data　is　averaged　to　match　the　long　time-scale　output　data.　Finally,　each　layer　of　the　improved　DFR　model　using　matching　data　fully　connects　the　parallel　and　cascade　forest　algorithms　to　improve　diversity　and　accuracy.　The　effectiveness　and　rationality　of　the　proposed　method　are　verified　on　the　real　DXN　dataset.　The　root　mean　square　error　of　the　proposed　method　is　17.4%,　14.0%,　and　13.1%　higher　than　those　of　the　no　filling,　linear　filling,　and　variational　autoencoder　filling　methods,　respectively.　IEEE