Workflow　scheduling　problems　have　been　widely　studied　in　cloud　computing　and　edge　computing,　which　aim　to　exploit　cloud-edge　resources　to　execute　workflow　tasks　considering　several　constraints　and　optimization　goals.　However,　in　the　era　of　Internet　of　things,　the　load　of　each　computing　task　and　the　amount　of　data　transferred　between　computing　tasks　will　fluctuate,　which　changes　the　original　workflow　and　needs　for　a　new　scheduling　plan　correspondingly.　Existing　methods　are　difficult　to　quickly　cope　with　these　dynamic　changes　and　there　are　few　studies　applying　neural　networks　to　solve　problems　in　workflow　scheduling.　To　bridge　the　gap,　we　propose　an　innovative　supervised　learning　method　which　leverages　function-fitting　strategy　of　neural　networks　to　link　the　workflow　environment　and　its　optimal　scheduling　plan.　Specifically,　our　approach　can　be　divided　into　two　steps,　the　first　one　is　to　generate　dataset　and　train　a　seq2seq-based　prediction　models.　In　this　step,　we　develop　an　algorithm　for　generating　a　significant　amount　of　workflow　instances　while　ensuring　dataset　diversity　based　on　complexity　estimation.　Then　we　apply　GA,　NSGA,　NSGA-NN　three　different　types　GA-based　optimization　methods　to　search　optimal　solutions.　Finally,　we　construct　dataset　which　includes　{workflow,　environment　configurations　-＞　obtained　optimal　solution}　and　train　a　seq2seq-based　model.　The　other　part　is　real-time　generation　of　scheduling　plans　based　on　trained　seq2seq-based　model.　Simulation　experiments　have　confirmed　that　our　method　is　both　effective　and　efficient,　demonstrating　its　ability　to　adapt　to　changes　in　the　execution　environment,　workflow　task　load,　and　task　data　transmission,　and　effectively　schedule　tasks　in　real-time.　The　simulation　results　show　that　the　seq2seq-based　prediction　method　can　approach　90%　of　the　optimal　scheme.