Modeling haloketones in drinking water using conformable neural networks: a case study of Jinhua, China

Abstract

The prediction of halogenated ketones in drinking water is relevant for public health surveillance and treatment control. Sixty-three samples from Jinhua, China, with routinely monitored physicochemical parameters were used, targeting three objectives: 1,1-dichloro-2-propanone (DCP), 1,1,1-trichloro-2-propanone (TCP), and total haloketones (HK). We compared two simple baselines—multiple linear regression and random forests—with an artificial neural network using radial basis function activation. The models were trained with a fixed training/validation/test split, minimum-maximum scaling to [0.1, 0.9], and evaluated with R, RMSE, and MAPE. A global sensitivity analysis identified the most influential inputs. The baselines established realistic performance limits (e.g., for DCP: R ≈ 0.77 and RMSE≈0.23 for linear regression; R ≈ 0.77 and RMSE≈0.28 for random forest). The conformable activation network improved agreement with observations for all targets: averaging R = 0.94, RMSE = 0.398. Sensitivity analysis was consistent with known factors in the process. The proposed activation design achieved strong gains over linear and tree-based baselines on a small dataset while remaining computationally light. We document the assumptions, data ranges, and limitations to support its reuse in routine monitoring.