Water　level　management　is　an　important　part　of　urban　water　system　management.　In　flood　season,　the　river　should　be　controlled　to　ensure　the　ecological　and　landscape　water　level.　In　non-flood　season,　the　water　level　should　be　lowered　to　ensure　smooth　drainage.　In　urban　areas,　the　response　of　the　river　water　level　to　rainfall　and　artificial　regulation　is　relatively　rapid　and　strong.　Therefore,　building　a　mathematical　model　to　forecast　the　short-term　trend　of　urban　river　water　levels　can　provide　a　scientific　basis　for　decision　makers　and　is　of　great　significance　for　the　management　of　urban　water　systems.　With　a　focus　on　the　high　uncertainty　of　urban　river　water　level　prediction,　a　real-time　rolling　forecast　method　for　the　short-term　water　levels　of　urban　internal　rivers　and　external　rivers　was　constructed,　based　on　long　short-term　memory　(LSTM).　Fuzhou　City,　China　was　used　as　the　research　area,　and　the　forecast　performance　of　LSTM　was　analyzed.　The　results　confirm　the　feasibility　of　LSTM　in　real-time　rolling　forecasting　of　water　levels.　The　absolute　errors　at　different　times　in　each　forecast　were　compared,　and　the　various　characteristics　and　causes　of　the　errors　in　the　forecast　process　were　analyzed.　The　forecast　performance　of　LSTM　under　different　rolling　intervals　and　different　forecast　periods　was　compared,　and　the　recommended　values　are　provided　as　a　reference　for　the　construction　of　local　operational　forecast　systems.