Coastal　water-quality　prediction　is　the　indispensable　work　to　prevent　the　red　tide　and　marine　pollution　accidents,　which　also　provides　the　effective　assistance　to　study　ocean　carbon　sink.　Due　to　the　multiple　inducing-factors　and　their　spatio-temporal　coupling　effects,　the　water-quality　prediction　not　only　needs　to　be　supported　by　big　data,　but　also　needs　an　effective　model　for　prediction　and　analysis.　However,　most　of　the　existing　models　frequently　use　timeline　data　from　the　same　section　or　local　collection　point,　and　cannot　realize　inversion　and　traceability　of　inducing-factors.　In　this　paper,　we　consider　these　tough　problems　and　propose　an　effective　neurodynamics-driven　prediction　model　for　state　evolution　of　coastal　water-quality　(NDPM-CWQ).　First,　an　event-driven　deep　belief　network　(EDBN)　is　designed　and　trained　using　the　spatio-temporal　data.　Second,　through　the　sensitivity　analysis　of　the　input　variables　in　EDBN　model,　we　rank　influence　degrees　of　spatio-temporal　variables　on　the　water-quality　and　give　the　inversion　and　traceability　of　inducing-factors.　Third,　the　convergence　of　training　EDBN　is　analyzed　from　the　perspective　of　the　stationary　distribution　and　decision　stability　of　Markov　chain.　Finally,　the　practical　data-based　experimental　results　show　that　the　proposed　NDPM-CWQ　not　only　achieves　better　prediction　performance,　but　also　can　quantitatively　analyze　the　inversion　and　traceability　of　inducing-factors.　IEEE