A Comparative Study of Multiple-Output and Single-Output Neural Network Models for Predicting Chemical Dosing in Water Treatment Systems

Accurate dosing of chemicals is essential for coagulation and disinfection in water treatment systems. Conventional jar testing is time-consuming and may not respond quickly to changing raw water quality. This study compares artificial neural network architectures for predicting polyaluminum chloride and chlorine dosages jointly using a multiple output (MO) model or separately using single output (SO) models. MO models may capture shared dosing patterns, while SO models allow target-specific specialization. Operational data from an industrial water treatment plant in Pathum Thani, Thailand were used as a case study. Twelve growing window experiments with grid search tuning were conducted under temporally ordered and seasonally varying conditions. Results show that MO and SO models achieved comparable accuracy, although SO models had slightly lower error variance. The framework may be adapted to other process control applications with related outputs. However, predictions represent historical dosing by operators rather than optimized chemical use.