With　the　rapid　development　of　economy,　the　problem　of　river　pollution　is　becoming　more　and　more　serious.　It　is　very　important　and　challenging　to　build　a　high-precision　water　quality　prediction　model　for　the　comprehensive　management　of　water　environment　and　prevention　of　water　pollution.　At　present,　some　achievements　have　been　made　in　data-driven　river　water　quality　prediction,　mainly　including　grey　model,　time　series　model,　support　vector　machine　model　and　neural　network　model.　However,　these　existing　prediction　methods　are　limited　to　single　station　prediction.　The　influence　of　the　upstream　water　quality　on　the　downstream　water　quality　is　ignored　and　the　upstream　stations　at　different　locations　will　have　different　effects　on　the　downstream　water　quality.　In　order　to　solve　the　problem　that　the　traditional　model　of　water　quality　prediction　does　not　consider　the　upstream　influence　and　difficulties　in　long-term　prediction,　a　new　water　quality　prediction　model　based　on　spatiotemporal　attention　mechanism　and　long-short-term　memory　neural　network　(TS-Attention-LSTM)　was　proposed,　which　considers　the　spatiotemporal　correlation　and　meteorological　factors.　Firstly,　the　spatial　attention　module　was　embedded　in　the　encoder　to　extract　the　significant　spatial　correlation　between　upstream　and　downstream.　The　interaction　among　the　water　quality　indicators　was　also　extracted.　Then,　the　temporal　attention　module　was　embedded　in　the　decoder　to　extract　the　important　time　series　features.　Moreover,　the　meteorological　factors　and　spatial　characteristics　were　fused　in　the　decoder.　Finally,　the　multi-step　prediction　of　water　quality　was　carried　out　by　using　LSTM　model.　In　this　paper,　the　research　area　located　in　Jinjiang　River　Watershed　in　Fujian　Province,　China.　The　nonpoint　source　pollution　in　this　river　basin　mainly　interrelated　with　livestock　discharge,　agricultural　field　runoff　and　rural　domestic　sewage.　The　results　show　that　the　TS-Attention-LSTM　model　can　effectively　capture　the　spatial　and　temporal　characteristics　of　water　quality　index　(such　as　dissolved　oxygen,　DO　and　total　phosphorus,　TP)　and　the　influence　of　meteorological　factors　in　Jinjiang　River　Basin.　The　mean　absolute　error　(MAE)　of　the　TS-Attention-LSTM　model　was　0.24,　the　mean　absolute　percent　error　(MAPE)　was　3.36%,　and　the　root　mean　square　error　(RMSE)　was　0.32,　which　performed　best　in　all　comparison　models.　The　determinable　coefficient　R2　was　0.5062,　performed　second　best　in　all　comparison　models.　©　2021　IEEE.