Numerical　models　are　of　fundamental　usage　for　estuarine　and　coastal　sciences.　Although　numerical　simulations　are　widely　applied,　analyzing　and　improving　them　are　often　challenging　tasks　given　their　large　volume　and　huge　parameter　space.　In　this　study,　a　novel　data-driven　framework　is　introduced　to　study　the　Minjiang　River　Plume　(MJRP).　The　framework　combines　Self-Organizing　Map　(SOM)　clustering　with　a　Hidden　Markov　Model　(HMM).　A　three-dimensional　Regional　Ocean　Model　System　for　MJRP　is　first　configurated　with　realistic　atmospheric,　oceanic,　and　riverine　forcings.　By　applying　SOM　clustering　to　the　modeled　sea　surface　salinity　(SSS)　with　similar　to　2,000　2-day　averaged　records　from　2010　to　2020,　we　identify　six　major　patterns　of　MJRP.　Each　pattern　exhibits　distinct　circulation　and　plume　structures.　These　MJRP　patterns　contain　not　only　seasonal　signals,　but　also　rich　short-term　variabilities　driven　by　the　riverine　inputs　and　oceanic　dynamics.　Then,　the　SOM-HMM　method　was　applied　to　predict　the　future　of　the　hidden　state　(i.e.,　patterns　of　MJRP)　from　the　observable　states　(wind　and　river　runoff).　With　a　hypothetic　SSS　product　from　a　geostationary　satellite　as　the　ground　truth,　we　show　that　the　SOM-HMM　method　can　predict　MJRP　patterns　considerably　high　prediction　accuracy　and　computational　efficiency.　Further,　these　patterns　were　translated　back　to　SSS　with　high　forecast　skills.　Combining　a　conventional　numerical　model　with　a　data-driven　method,　this　approach　can　be　promisingly　applied　in　the　short-term　marine　forecast　to　support　the　utilization　and　management　of　other　estuaries.