|
Comparative analysis of MO and SO ANN for integrated chemical dosing in water treatment plants |
|---|---|
| รหัสดีโอไอ | |
| Title | Comparative analysis of MO and SO ANN for integrated chemical dosing in water treatment plants |
| Creator | Punsita Bualert |
| Contributor | Warut Pannakkong, Advisor |
| Publisher | Thammasat University |
| Publication Year | 2568 |
| Keyword | Artificial neural network, Multiple-output model, Chemical dosing prediction, Growing window validation, Water treatment plant |
| Abstract | Accurate chemical dosing in water treatment plants (WTPs) remains challenging due to the nonlinear and seasonally variable relationship between raw water quality and the optimal dosages of coagulant and disinfectant. This study developed and compared Multiple-Output (MO) and Single-Output (SO) Multilayer Perceptron (MLP) Artificial Neural Network (ANN) models for the simultaneous and independent prediction of Polyaluminium Chloride (PAC) and liquid chlorine (Cl) dosages at an industrial WTP in Pathum Thani, Thailand. A dataset of 1,089 daily operational records from January 2019 to December 2021 was used, with input features comprising raw water quality parameters, production water intake, dynamic quality change variables incorporating operator-defined treated water targets, and temporal encoding features. Model performance was evaluated across twelve growing-window repeated holdout experiments covering all twelve months of 2021, with hyperparameters selected through grid search optimization. The Wilcoxon signed-rank test was applied to assess statistical significance of architecture differences. For PAC prediction, the SO architecture achieved a lower average RMSE of 96.95 kg against 109.99 kg for the MO model across twelve experiments, though this difference was not statistically significant (W = 20, p = 0.1514). For liquid Cl prediction, the SO architecture demonstrated a statistically significant advantage with an average RMSE of 22.49 kg compared to 39.80 kg for the MO model (W = 3, p = 0.0024). The MO architecture showed competitive performance during extreme La Nina monsoon conditions, suggesting complementary deployment potential under high climate variability |
