Scholarlink

<p>Accurate short-term prediction of key water quality (WQ) parameters is essential for improving treatment efficiency and operational decision-making in municipal water treatment plants. This study investigates the performance of standalone, ensemble, and hybrid machine learning (ML) models for predicting hardness (HD) and turbidity (Turb) at the Tamburawa Water Treatment Plant (TWTP), Kano State, Nigeria. A total of 209 observations were used to develop Fine Tree (FT), Bagged Tree (BT), Gaussian Process Regression (GPR), linear regression, single neural network, and hybrid ensemble–neural network models (FT–NN, BT–NN, and GPR–NN). Two predictor configurations were examined: COMBO I, comprising ionic parameters Calcium (Ca), alkalinity (Alk), total dissolved solids (TSS), and Free CO<sub>2</sub>, and COMBO II, consisting of routinely monitored physicochemical variables water temperature (WT), turbidity (Turb), pH, and electrical conductivity (EC). Model evaluation was performed using tenfold cross-validation, with performance assessed via normalized RMSE, MSE, and MAE. Results show that hybrid models consistently outperform standalone and baseline approaches. For HD prediction, the BT–NN model achieved RMSE and MAE values of 0.095 and 0.065 under COMBO I, and 0.101 and 0.069 under COMBO II, representing approximately 12–18% error reduction relative to standalone ensemble learners. For Turb prediction, BT–NN produced RMSE values of 0.111 (COMBO I) and 0.108 (COMBO II), outperforming linear regression and single neural network baselines by over 20%. Feature importance analysis identifies calcium and alkalinity as dominant drivers of HD under COMBO I, while electrical conductivity and pH were most influential under COMBO II; for Turb, lagged turbidity and electrical conductivity were the primary predictors. Sensitivity analysis confirmed the robustness of the hybrid models to ± 10% input perturbations. Overall, the findings demonstrate that hybrid ensemble neural network models provide accurate, interpretable, and computationally efficient tools for short-term prediction of HD and Turb using readily measurable parameters. The study establishes methodological feasibility and supports the use of hybrid ML models as decision-support systems in resource-constrained water treatment environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hybrid neural network ensembles for predicting water quality in municipal treatment systems

  • 期刊论文

  • 发表日期：

    2026年2月2日

查看全文
  • I. A. Mahmoud,
  • M. Alsubih,
  • M. Alkahtani,
  • S. Sa’id,
  • U. U. Aliyu

摘要

Accurate short-term prediction of key water quality (WQ) parameters is essential for improving treatment efficiency and operational decision-making in municipal water treatment plants. This study investigates the performance of standalone, ensemble, and hybrid machine learning (ML) models for predicting hardness (HD) and turbidity (Turb) at the Tamburawa Water Treatment Plant (TWTP), Kano State, Nigeria. A total of 209 observations were used to develop Fine Tree (FT), Bagged Tree (BT), Gaussian Process Regression (GPR), linear regression, single neural network, and hybrid ensemble–neural network models (FT–NN, BT–NN, and GPR–NN). Two predictor configurations were examined: COMBO I, comprising ionic parameters Calcium (Ca), alkalinity (Alk), total dissolved solids (TSS), and Free CO2, and COMBO II, consisting of routinely monitored physicochemical variables water temperature (WT), turbidity (Turb), pH, and electrical conductivity (EC). Model evaluation was performed using tenfold cross-validation, with performance assessed via normalized RMSE, MSE, and MAE. Results show that hybrid models consistently outperform standalone and baseline approaches. For HD prediction, the BT–NN model achieved RMSE and MAE values of 0.095 and 0.065 under COMBO I, and 0.101 and 0.069 under COMBO II, representing approximately 12–18% error reduction relative to standalone ensemble learners. For Turb prediction, BT–NN produced RMSE values of 0.111 (COMBO I) and 0.108 (COMBO II), outperforming linear regression and single neural network baselines by over 20%. Feature importance analysis identifies calcium and alkalinity as dominant drivers of HD under COMBO I, while electrical conductivity and pH were most influential under COMBO II; for Turb, lagged turbidity and electrical conductivity were the primary predictors. Sensitivity analysis confirmed the robustness of the hybrid models to ± 10% input perturbations. Overall, the findings demonstrate that hybrid ensemble neural network models provide accurate, interpretable, and computationally efficient tools for short-term prediction of HD and Turb using readily measurable parameters. The study establishes methodological feasibility and supports the use of hybrid ML models as decision-support systems in resource-constrained water treatment environments.