The　prediction　of　accurate　water　quality　has　great　significance　to　the　sustainable　management　of　water　resources　and　pollution　prevention.　Due　to　the　complexity　of　water　environment,　it　is　difficult　to　do　so.　Traditional　prediction　methods　are　mainly　linear　methods.　Their　prediction　accuracy　is　limited　since　they　fail　to　reflect　nonlinear　characteristics　in　water　quality　data.　To　achieve　much　higher　accuracy,　this　work　proposes　to　combines　a　Savitzky-Golay　filter　with　Attention-based　Long　Short-Term　Memory　to　perform　a　multi-step　prediction　of　water　quality.　The　proposed　model　uses　a　Savitzky-Golay　filter　for　smoothing　sequences　to　reduce　noise　interference.　The　adoption　of　an　attention　mechanism　can　extract　effective　information　from　complex,　long,　and　temporal　dependence.　Experimental　results　demonstrate　that　the　proposed　method　outperforms　other　state-of-the-art　peers.　©　2020　IEEE.