The　prediction　of　water　quality　has　great　significance　for　the　management　of　water　environment　and　the　protection　of　water　resources.　Traditional　water　quality　prediction　methods　are　relatively　simple,　and　linear　models　are　often　used　to　predict　water　quality.　However,　such　models　limit　the　accuracy　of　prediction　and　lack　the　analysis　of　nonlinear　characteristics　of　water　quality.　In　addition,　due　to　the　complex　water　environment,　the　water　quality　time　series　has　large　noise,　which　makes　it　difficult　for　traditional　models　to　effectively　predict　water　quality　indicators　under　complex　environmental　conditions.　To　solve　this　problem,　this　work　proposes　an　integrated　prediction　method　that　combines　Savitzky-Golay　filter　with　Long　Short-Term　Memory　(LSTM)-based　Encoder-Decoder　neural　network　to　predict　water　quality　at　the　next　time　interval.　In　this　approach,　the　water　quality　time　series　is　first　smoothed　by　Savitzky-Golay　filter,　and　LSTM　can　extract　valid　information　from　complex　time　series.　Based　on　them,　an　integrated　model　is　for　the　first　time　established　and　can　well　characterize　statistical　characteristics.　Experimental　results　demonstrate　that　it　achieves　better　prediction　results　than　some　typical　prediction　methods.　©　2019　IEEE.