Experimental Comparison of Improved GBDT Methods for Water Quality Parameter Prediction

As an important environmental resource, water is significant for maintaining human healthy living and promoting national economic development. Accurately predicting the pH values and conductivities, which are reliable indicators of water quality, can help respond in advance to abnormal water quality conditions. To improve the accuracy of water quality prediction, three improved gradient boosting decision tree (GBDT) methods, namely Categorical Boosting (CatBoost), Light Gradient Boosting Machine (LightGBM), and eXtreme Gradient Boosting (XGBoost), were employed to predict the pH values and conductivities and compared with the Support Vector Machine (SVM), Random Forest (RF), and Long Short-Term Memory (LSTM) models. Using the physicochemical monitoring dataset of Taihu Lake from 2007 to 2015, extensive experiments were conducted. The experimental results were computed as statistics and further analyzed using the SHAP method. The statistical results show that all three methods achieved high prediction accuracies. The prediction accuracies of CatBoost and XGBoost were generally higher than those obtained by LightGBM and also outperformed the SVM, RF, and LSTM methods.