ScholarWorks@동국대학교

초록

Ensuring a safe and sustainable drinking water supply is challenging due to raw water quality fluctuations, complex treatment processes, and stringent regulatory requirements. Accurately predicting chemical dosages and treatment outcomes is essential for optimizing treatment efficiencies. Although deep learning models have demonstrated strong predictive performance, their application in water treatment remains limited by insufficient feature-level interpretability for operational decision-making. This study introduces a novel transformer-based framework incorporating a dual-level feature attention mechanism (i.e., single-feature attention for temporal patterns per feature and cross-feature attention) to enhance interpretability while achieving high predictive accuracy. The proposed model was applied to predict chemical dosages (coagulant [mg L−1], pre- and mid-chlorine [mg L−1]) and settled water quality (turbidity [NTU] and residual chlorine [mg L−1]) in a water treatment plant. Compared with baseline models (MLP, GRU, CNN-GRU, and Autoformer), the proposed model consistently improved prediction accuracy (R2) by 2−232% and reduced error metrics (MAE, MAPE, MSE, and nRMSE) by 3−98% across all target variables. Feature importance scores derived from the proposed model revealed that raw water turbidity was the dominant predictor of coagulant dosages and settled water turbidity, while raw water TOC strongly influenced chemical dosages and residual chlorine. By providing accurate predictions together with transparent feature-level interpretability, the proposed framework supports data-driven chemical dosing strategies, contributing to chemical savings and ultimately to cleaner water treatment operations. © 2026 Elsevier Ltd.

키워드