Method for establishing predictive models for total organic halogen based on piecewise interpolation and machine learning

In disinfection by-product (DBP) research, the parameter ‘total organic halogen’ (TOX) is a significant aggregate indicator and reports the total content of halogenated DBP in water, determined in a single experimental process. TOX modeling can facilitate the prediction, diagnosis, and control of the drinking water disinfection process. The modeling approach is often based on the reaction mechanisms of the disinfection process. However, building an accurate TOX model is difficult due to the complexity and nonlinearity of the disinfection reaction mechanisms, and many simplifications have been made in the modeling process, resulting in poor adaptability of the TOX model in practical applications. Machine learning algorithms are data-driven modeling methods that can achieve high prediction accuracy and are simple and convenient to apply. However, in practical experiments, the TOX dataset is often small (usually < 10 points), making TOX modeling through machine learning algorithms particularly difficult. To solve this issue, this study established a method using piecewise interpolation to expand the TOX dataset and subsequently machine learning algorithms to establish the model. Three common machine learning algorithms, backpropagation neural network, radial basis function neural network, and support vector machine, were used to evaluate the data expansion method. The modeling of TOX for a chloramination and chlorination disinfection process shows that this method can achieve satisfactory results regarding sensitivity and accuracy. All the models provided favorable predictions, with relatively high correlation coefficients (> 0.99) and low mean square errors (< 5.31 × 10−5). [Display omitted] •This study established a method to predict total organic halogen after disinfection with machine learning algorithms.•The problem of insufficient data in total organic halogen prediction with machine learning was solved by dataset expansion.•After expanding of the dataset, satisfactory prediction were obtained through commonly used machine learning algorithms.•The prediction results obtained with machine learning algorithms were close to the results of mechanism models.