Surface　water　quality　is　increasingly　deteriorated　in　recent　years,　and　therefore,　high-quality　early　warning　and　prediction　of　water　quality　are　essential　for　sustainability　of　water　resources　and　emergency　response　mechanisms.　Long　short-term　memory　(LSTM)　network　is　widely　applied　in　the　existing　literature　on　the　prediction　of　water　quality　time　series.　However,　only　applying　LSTM　for　the　prediction　of　water　quality　time　series　cannot　well　address　irregular　fluctuations　in　the　water　quality　series　caused　by　multiple　complex　factors.　To　solve　this　problem,　a　data-driven　prediction　model　for　the　water　quality　time　series　was　proposed,　named　STL-LSTM-ED,　which　was　composed　of　seasonal-trend　decomposition　using　locally　weighted　scatterplot　smoothing　(STL)　and　LSTM　based　on　encoder-decoder　(LSTM-ED).　Compared　with　several　typical　models　of　LSTM,　LSTM-ED,　and　a　sequence　decomposition　method　based　on　LSTM,　the　proposed　STL-LSTM-ED　can　significantly　improve　the　prediction　accuracy　and　reliability　of　the　water　quality　time　series,　and　also　provide　the　effective　data　support　for　dynamic　warning　of　water　quality.　©　2022　Chinese　Journal　of　Intelligent　Science　and　Technology.　All　Rights　Reserved.