This study focuses on predicting water quality in urban rivers, specifically the Fenjiang River in Foshan City, using wavelet packet denoising (WPD) and long short-term memory (LSTM) models. The research aims to improve the accuracy of water quality predictions by reducing noise in historical monitoring data and enhancing the performance of LSTM models. The study collected and analyzed water quality data from four monitoring stations along the Fenjiang River from 2016 to 2021. The main pollutants of interest were chemical oxygen demand (COD) and ammonia nitrogen (NH₃-N). WPD was applied to denoise the raw data, and an LSTM model was developed to predict these pollutants. The results show that the WPD-LSTM model outperformed traditional LSTM models in terms of predictive accuracy. Specifically, the RMSE values for NH₃-N predictions decreased by 55% to 67% (average decrease of 61%) and for COD predictions by 18% to 51% (average decrease of 29%) over a 12-hour prediction period. Over a 3-day prediction period, the RMSE values for NH₃-N decreased by 40% to 83% (average decrease of 65%), and for COD by 50% to 69% (average decrease of 60%). The study concludes that the WPD-LSTM model significantly improves the accuracy of water quality predictions, providing valuable insights for effective water environment management in urban river systems. The model's enhanced performance can help water environment departments make more informed decisions and improve water quality monitoring and management.This study focuses on predicting water quality in urban rivers, specifically the Fenjiang River in Foshan City, using wavelet packet denoising (WPD) and long short-term memory (LSTM) models. The research aims to improve the accuracy of water quality predictions by reducing noise in historical monitoring data and enhancing the performance of LSTM models. The study collected and analyzed water quality data from four monitoring stations along the Fenjiang River from 2016 to 2021. The main pollutants of interest were chemical oxygen demand (COD) and ammonia nitrogen (NH₃-N). WPD was applied to denoise the raw data, and an LSTM model was developed to predict these pollutants. The results show that the WPD-LSTM model outperformed traditional LSTM models in terms of predictive accuracy. Specifically, the RMSE values for NH₃-N predictions decreased by 55% to 67% (average decrease of 61%) and for COD predictions by 18% to 51% (average decrease of 29%) over a 12-hour prediction period. Over a 3-day prediction period, the RMSE values for NH₃-N decreased by 40% to 83% (average decrease of 65%), and for COD by 50% to 69% (average decrease of 60%). The study concludes that the WPD-LSTM model significantly improves the accuracy of water quality predictions, providing valuable insights for effective water environment management in urban river systems. The model's enhanced performance can help water environment departments make more informed decisions and improve water quality monitoring and management.