The use of explainable artificial intelligence for interpreting the effect of flow phase and hysteresis on turbidity prediction

Predicting turbidity (T), which represents the amount of fine sediment in water, is essential in effective water quality management. In this study, two ensemble learning models, XGBoost and light gradient boosting decision tree (LGB), were employed to predict T, using discharge (Q) as an independent variable. The input variables were classified into three groups based on the flow phase: rising limb, falling limb, and base flow, where different time-frequency datasets (2, 8, and 24 h) were utilized to develop the model. In the first model set (Model 1), each model was trained separately for every phase, and their performance was tested by applying each to the corresponding Q. Another model set using XGBoost and LGB was developed by considering the entire period without classification for a comparison purpose (Model 2). The results demonstrated that Model 1 which used data classified into three phases outperformed Model 2. Further analysis of the flood phase and hysteresis in the relationship between Q and T showed that different data distributions in the three phases determined the performance differences between Models 1 and 2. By considering these differences, Model 1 exhibited better performance compared to Model 2. The Shapley additive explanation (SHAP), a novel explainable artificial intelligence method, provided a reasonable interpretation of the difference in model predictions between Models 1 and 2.

Automated summary

This study employed two ensemble learning models, XGBoost and LGB, to predict turbidity (T) in water, which is crucial for effective water quality management. The input variables were classified into three groups based on the flow phase, and different time-frequency datasets were utilized to develop the models. The results showed that the model (Model 1) using data classified into three phases outperformed the model without classification (Model 2). Further analysis revealed that the performance differences between Models 1 and 2 were determined by the different data distributions in the three phases. By considering these differences, Model 1 exhibited better performance compared to Model 2. The Shapley additive explanation (SHAP) provided a reasonable interpretation of the difference in model predictions between Models 1 and 2.