Liang Youheng, Huang Ruixing, Wang Jingrui, Han Zhengpeng, Wu Sisi, Tan Yao, Huangfu Xiaoliu, He Qiang
Key Laboratory of Eco-Environments in Three Gorges Reservoir Region, Ministry of Education, College of Environment, and Ecology, Chongqing University, Chongqing 400044, PR China.
State Key Laboratory of Urban Water Resources and Environment, School of Municipal and Environmental Engineering, Harbin Institute of Technology, Harbin 150090, PR China.
Water Res. 2025 Sep 1;283:123849. doi: 10.1016/j.watres.2025.123849. Epub 2025 May 17.
Chlorination and chloramination as common water disinfection methods are challenged by the unintended formations of hazardous disinfection by-products (DBPs). Accurately predicting DBP formation is essential for improving water treatment processes and protecting public health. However, existing models for predicting DBP levels in drinking water treatment, especially for unregulated DBPs, are insufficient. In this study, we developed machine learning (ML) models to predict the levels of five total DBPs (TDBPs) and their ten individual DBPs (IDBPs) resulting from chlorination and chloramination, covering both regulated and unregulated DBPs. To solve the challenge of redundant models, we adopted a data integration strategy to construct larger-scale unified models. The results suggested that the unified model performance outperformed individual models, whereas the individual models were more effective for predicting TDBPs. Moreover, the Shapley additivity interpretation and partial dependence plots provided valuable insights into the key factors influencing DBP formation, aligning with experimental findings. A web application, known as the ACAI platform, was deployed for the first time to predict DBP levels using an automated ML protocol. This user-friendly platform makes DBP prediction accessible to a wide range of users, including those without programming expertise. We expect that these ML models and web interface will support data-driven decision-making in disinfection.
氯化和氯胺化作为常见的水消毒方法，面临着意外形成有害消毒副产物（DBP）的挑战。准确预测消毒副产物的形成对于改进水处理工艺和保护公众健康至关重要。然而，现有的用于预测饮用水处理中消毒副产物水平的模型，尤其是针对未受监管的消毒副产物的模型，并不完善。在本研究中，我们开发了机器学习（ML）模型来预测氯化和氯胺化产生的五种总消毒副产物（TDBP）及其十种单个消毒副产物（IDBP）的水平，涵盖了受监管和未受监管的消毒副产物。为了解决冗余模型的挑战，我们采用了数据集成策略来构建更大规模的统一模型。结果表明，统一模型的性能优于单个模型，而单个模型在预测总消毒副产物方面更有效。此外，夏普利加性解释和部分依赖图为影响消毒副产物形成的关键因素提供了有价值的见解，与实验结果一致。首次部署了一个名为ACAI平台的网络应用程序，使用自动化机器学习协议来预测消毒副产物水平。这个用户友好的平台使广泛的用户，包括那些没有编程专业知识的用户，都能够进行消毒副产物预测。我们期望这些机器学习模型和网络界面将支持消毒过程中数据驱动的决策。
