Coupling generative and predictive machine learning algorithms to enhance haloacetonitriles prediction in small water systems.

To overcome data scarcity constraints in modeling emerging disinfection byproducts (EDBPs) for resource-limited small water systems, this study introduces a novel framework coupling generative algorithms (GAs) and machine learning algorithms (MLAs). We evaluated models integrating three GAs (CWGAN-GP, VAE, and Add Noise) and four MLAs (GRNN, XGBoost, GPR, and SVR) using full-scale sampling data. Data quality analysis confirmed that the Conditional Wasserstein Generative Adversarial Network with Gradient Penalty (CWGAN-GP) generated data with the highest consistency with original samples. Sensitivity analysis indicated that model performance improved with the volume of generated data up to an optimal threshold. In classification of dibromoacetonitrile exceedance, all coupled models showed significant improvements in accuracy, precision, and specificity over baseline MLAs. The optimal coupling, CWGAN-GP + Generalized Regression Neural Network (GRNN), increased classification accuracy from 74.1% to 81.5% and specificity from 45.8% to 62.5%. For dichloroacetonitrile concentration prediction, CWGAN-GP significantly enhanced GRNN (MSE decreased from 0.152 to 0.105) and eXtreme Gradient Boosting (XGBoost) (MSE decreased by 31.3%); while Variational Autoencoder (VAE) greatly improved Gaussian Process Regression (GPR) (MSE decreased by 16.0%) and Support Vector Regression (SVR) (MSE decreased by 14.2%). This GA-MLA framework reduces reliance on costly sampling, lowering modeling expenses by up to 49.7%. This coupled framework effectively mitigates data limitations, providing a viable and economical pathway for EDBP surveillance in small water systems.