Data-driven prediction of neutralizer pH and valve position towards precise control of chemical dosage in a wastewater treatment plant

Abstract

Proper pH control is important to the neutralization of industrial wastewater that will facilitate downstream biological treatment and can strongly affect the utilization of chemical reagent/resource. This study aimed to apply machine learning (ML) models for both pH prediction and lime dosage control towards enhanced automated control of neutralization processes. To achieve this goal, eight ML models were employed and compared in modeling performance, and optimized by using correlation analysis, cross-validation, and grid search techniques. In the neutralizer pH prediction, the highest coefficients of determination (R2) results were obtained at 0.765 (k-nearest neighbors - KNN), 0.918 (eXtreme Gradient Boosting - XGBoost), and 0.900 (random forest - RF) for three neutralizer tanks, accompanied with the lowest root-mean-square error (RMSE) values of 0.289, 0.100, and 0.093, respectively. The impacts of input features were quantified by sensitivity analysis using SHAP values, which demonstrated the importance of temperature, valve position, and upstream pH as high as 0.214, 0.156, and 0.118 (the mean of absolute SHAP value). For lime dosage control, the best model performance came from XGBoost (R2 values of 0.605) for valve 1, RF (0.788) for valve 2, and RF (0.436) for valve 3 with the corresponding RMSE values of 8.056, 6.125, and 4.466, respectively. The recommended valve position was based on the target pH and some examples were illustrated at different upstream pH values. The results of this study has demonstrated that ML approach can be an effective tool to help conserve chemical resources via enhanced chemical dosage control in wastewater treatment.