Extremely randomized tree: a new machines learning method for predicting coagulant dosage in drinking water treatment plant

Coagulation using metal salts such as aluminum sulfate and ferric sulfate is the most well-known method used during the coagulation–flocculation process and mainly adopted in the drinking water treatment plants worldwide. The most method for determining the optimal coagulant dosage is the jar test, but this is clearly laborious and time-consuming task. Widely used regression models such as multiple linear regression (MLR) are unable to provide a high linking between water quality variables and the optimal coagulant dosage, due to the high nonlinearity and the multiple factors affecting the coagulation process. In this chapter, we propose a new robust method for predicting coagulant dosage using machine learning approaches. We proposed and compared two methods, namely, extremely randomized tree (ERT) and random forest (RF) models. To demonstrate the usefulness and robustness of the proposed models, a result using the MLR models was also provided for further comparison. The models were developed using several water quality variables selected as input of the models, namely, turbidity, pH, dissolved oxygen, electrical conductivity, and the water temperature. The accuracy of the models was evaluated using coefficient of correlation (R), Nash–Sutcliffe efficiency (NSE), root-mean-square error, and mean absolute error. According to the obtained results, the ERT approach was robust across methods tested. Both ERT and RF provided high accuracy during the training and validation phases; however, standard MLR model showed a lowest accuracy both in training and validation phases and the coagulant dosage was often incorrectly predicted. During the validation phase, the R and NSE values of the ERT model were the greatest with values equal to 0.899 and 0.790, respectively, greater than the values provided by the RF model (R=0.851, NSE=0.708), and significantly higher than the values provided by the MLR model, indicating that the ERT was the best and supporting the use of this model for predicting coagulant dosage in drinking water treatment plant.