Biochar adsorption technology has been widely used to remove ammonia nitrogen from water bodies. However, existing methods for predicting adsorption efficiency often lack sufficient accuracy and practical usability. This study evaluated eight machine learning models, including XGB, LR, KNN, DT, RF, GBR, SVR, and ANN, to predict the adsorption efficiency of ammonia nitrogen. The evaluation utilized a dataset comprising 770 instances of ammonia nitrogen adsorption by biochar. The models' prediction performances were systematically compared, and cross-validation was applied to enhance their generalization ability, leading to the selection of the best-performing model. The selected model's parameters were further optimized using Bayesian optimization to improve the prediction accuracy. The Bayesian-optimized XGB model achieved the highest predictive performance, with a coefficient of determination (R2) of 0.978. The R2 values of the other models ranged from 0.556 (LR) to 0.927 (RF). Key factors influencing ammonia nitrogen adsorption efficiency were identified using SHAP analysis. These factors included the amount of biochar, adsorption time, initial ammonia nitrogen concentration, solution pH, pyrolysis time, and O%. Their optimal ranges were further determined through partial dependency plots. This study developed a reliable machine learning tool for accurately predicting ammonia nitrogen adsorption efficiency. Additionally, it provided insights into optimizing the preparation processes and adsorption conditions of biochar, contributing to its practical application in treating ammonia nitrogen pollution in water bodies.
Read full abstract