Nowadays,　the　applications　of　water　quality　prediction　in　the　field　of　regional　water　environment　management　are　increasing.　It　refers　to　predicting　the　elemental　values　of　the　water　environment　in　the　future　based　on　past　monitoring　data,　which　is　essential　to　realize　the　real-time　evaluation　of　water　quality　and　dynamic　control　of　pollution　sources.　However,　the　water　environment　indicators　are　affected　by　various　elements,　which　have　a　large　volatility　and　non-linear　characteristics.　In　addition,　most　of　the　existing　water　quality　predictions　focus　on　single-step　predictive　modeling　of　single　elements　of　the　water　environment　and　lack　multi-step　predictive　analysis　of　multifactor　data　of　the　water　environment.　In　this　paper,　a　novel　long-term　prediction　model　based　on　genetic　simulated　annealing-based　particle　swarm　optimization　(GSPSO)　with　seasonal-trend　decomposition　using　LOESS　(STL)　is　proposed　and　named　GSPSO-STL-Autoformer　(GS-Autoformer).　It　realizes　the　multi-factor　and　long-term　prediction　of　water　quality　time　series　data.　Firstly,　the　Autoformer＇s　hyperparameters　are　optimized　by　the　GSPSO　to　improve　its　convergence　speed.　Secondly,　the　multi-factor　features　are　decomposed　by　the　STL　to　make　the　model　more　focused　on　learning　feature　information　of　each　component.　Finally,　the　long-term　prediction　is　realized　by　the　Autoformer.　Comparative　experiments　with　state-of-the-art　peers　show　that　the　GS-Autoformer　can　effectively　improve　the　accuracy　of　multi-factor　and　long-term　predictions.　©　2024　IEEE.