Prediction of Total Phosphorus Based on Distance Correlation and Machine Learning Methods-a Case Study of Dongjiang River, China

Elevated total phosphorus (TP) leads to water eutrophication, affecting aquatic life and ecosystems. Effective control of phosphorus concentration in rivers is vital. Detecting TP and other nutrients is essential for evaluating river health. Establishing real-time TP prediction is key for studying eutrophication and blooms. In this study, we compared the prediction models combined with two different feature selection tools, Pearson and distance correlation coefficient, and evaluated them with four evaluation indicators. It was determined that the assessment outcomes of the distance correlation coefficient yielded heightened precision. Notably, the random forest validation contributed to an impressive 8.3% enhancement in the performance R2 of the model, and the utilization of BP neural network validation resulted in an improvement of 2.7% in the model's performance. A real-time TP prediction model using distance correlation coefficient was eventually created, which reaches an accuracy up to 82.6%. The final model was applied to field data, revealing elevated TP levels downstream of Dongjiang River in China due to high population density and numerous factories. Additionally, the results indicated that the distance correlation coefficient outperformed the Pearson correlation coefficient in predicting the total phosphorus in Dongjiang River. Our findings suggest that combining distance correlation coefficients with machine learning holds promise for more accurate water quality prediction models. These findings establish a foundation for enhancing the efficacy of the water quality prediction model.