Water　quality　is　a　basic　work　in　environmental　governance,　which　has　vital　significance　in　promoting　the　sustainable　utilization　of　water　resources　and　instant　pollution　prevention　and　precise　control.　Water　quality　data　is　dynamic　and　frequently　fluctuating　with　different　temporal　and　spatial　dimensions,　therefore　it　can　be　challenging　to　predict.　A　hybrid　AM-ConvLSTM　deep　learning　algorithm　is　proposed　in　this　paper　to　rapidly　predict　the　trend　of　water　quality　which　can　run　faster　and　require　low　computing　power　rather　than　the　traditional　MIKE　21　hydrological　method.　The　ConvLSTM　method　and　the　attention　mechanism　are　assembled　to　build　AM-ConvLSTM　model　to　better　capture　spatial　correlation.　Moreover,　the　statistic　methods　are　used　to　evaluate　the　effectiveness　of　the　model　and　then　compared　with　varieties　of　deep　learning　baseline　methods.　The　results　reveal　that　the　hybrid　AM-ConvLSTM　model　can　effectively　replace　MIKE　21　model　to　predict　the　future　trend　of　water　quality,　and　then　the　local　environmental　protection　agencies　will　respond　quickly　to　emergency　water　pollution.　©　2021　IEEE.