Application of Machine Learning in Water Quality Prediction and Analysis for River Cross-Sections
Water quality prediction is essential for river basin management, yet existing models often struggle with non-stationary, noisy monitoring data. This study collected water quality data from two city-level control sections in southern China from December 2020 to June 2024, including eight indicators: water temperature, turbidity, pH, conductivity, dissolved oxygen (DO), ammonia nitrogen (NH4+-N), total phosphorus (TP), and permanganate index (CODMn). To predict four key indicators (DO, NH4+-N, TP, CODMn), we developed hybrid models combining seasonal trend decomposition (STD), Bayesian hyperparameter optimization, and either random forest (RF) or XGBoost. STD smoothed and denoised the data while extracting seasonal factors; Bayesian optimization tuned model hyperparameters. Evaluation showed that the STD-Bayesian-XGBoost model achieved smaller bias errors and higher prediction accuracy than STD-Bayesian-RF. Specifically, XGBoost reduced root mean square error (RMSE) by 15-20% across all four indicators and improved the coefficient of determination (R²) to above 0.90, compared to RF's 0.85-0.88. The models were validated on southern river data, but the methodology is generalizable to other climatic and hydrological settings. This work provides a technical reference for pollution reduction and carbon management in regional watersheds.