Prediction of Eutrophic Water Quality in the Daluxi River Based on a Multi-Scale Feature Extraction and Hybrid Screening Strategy

As water environmental issues in watersheds shift from traditional pollution to eutrophication and degradation of aquatic ecosystems, accurately predicting water quality trends has become crucial for eutrophication control. This study focuses on the Daluxi sub-watershed, a primary tributary of the Yangtze River, to develop a model for predicting eutrophication trends. Using wavelet analysis and Pearson correlation analysis, we explored the interactions between meteorological and water quality factors, as well as the influence of upstream and downstream water quality factors, ensuring the explanatory power of the input variables and laying a foundation for the model's predictive accuracy and robustness. Through a comparative analysis of various machine learning models, Gradient Boosting Decision Tree (GBDT) was ultimately selected as the optimal predictive model. Feature importance and Partial Dependence Plot (PDP) analyses were employed to quantify the contribution of each influencing factor to the prediction results. The findings show that the model performs well in predicting downstream water quality indices such as TN, COD _Mn , AN, and TP for the following day, with R² values of 0.90, 0.89, 0.88, and 0.89, respectively; Mean Absolute Error (MAE) values of 0.14, 0.27, 0.03, and 0.01, respectively; Root Mean Squared Error (RMSE) values of 0.22, 0.37, 0.06, and 0.01, respectively, demonstrating excellent predictive performance. This study enables early warnings of downstream water quality based on upstream water quality, facilitating the timely identification of potential water quality issues and providing a scientific basis for timely control measures, thereby effectively supporting water resource management and eutrophication control efforts.