Scholarlink

<p>Accurate prediction of the Langelier Saturation Index (LSI) is essential for assessing water stability, particularly in managing corrosion and scaling in water distribution systems. Traditional estimation methods often rely on parameters that are difficult to measure in situ, limiting their utility for real-time monitoring. This study proposes an advanced deep learning framework incorporating Long Short-Term Memory (LSTM), Clockwork Recurrent Neural Networks (CWRNN), and two hybrid extensions: CWRNN augmented with a Memory Network (CWRNN-MN) and a further optimized version using the Teaching–Learning-Based Optimization algorithm (TLBO-CWRNN-MN). Models were trained on pH, temperature, and total dissolved solids to enable rapid, cost-effective prediction of LSI without the need for complex laboratory analyses. While the LSTM model exhibited limited accuracy in early scenarios, its performance improved substantially with the inclusion of extended temporal inputs. Nonetheless, the hybrid architectures significantly outperformed the baseline models. In the most comprehensive scenario, TLBO-CWRNN-MN achieved the highest predictive performance (<i>R</i> = 0.982, RMSE = 0.054, NRMSE = 0.034), followed by CWRNN-MN (<i>R</i> = 0.972, RMSE = 0.079, NRMSE = 0.049), underscoring the effectiveness of integrating memory components and metaheuristic optimization in capturing nonlinear temporal dependencies. These findings are especially pertinent to regions such as West Azerbaijan Province, where escalating water scarcity and groundwater depletion have intensified ecological vulnerabilities, notably the decline of Lake Urmia. The proposed models offer a scalable and reliable approach to predictive water quality management, providing critical support for infrastructure sustainability and long-term environmental stewardship.</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

Smart Water Quality Monitoring Using a Hybrid Deep Learning Framework with Memory Network and Teaching–Learning-Based Optimization for Accurate LSI Prediction

  • 期刊论文

  • 发表日期：

    2025年8月11日

查看全文
  • Milad Sharafi,
  • Javad Behmanesh,
  • Vahid Rezaverdinejad

摘要

Accurate prediction of the Langelier Saturation Index (LSI) is essential for assessing water stability, particularly in managing corrosion and scaling in water distribution systems. Traditional estimation methods often rely on parameters that are difficult to measure in situ, limiting their utility for real-time monitoring. This study proposes an advanced deep learning framework incorporating Long Short-Term Memory (LSTM), Clockwork Recurrent Neural Networks (CWRNN), and two hybrid extensions: CWRNN augmented with a Memory Network (CWRNN-MN) and a further optimized version using the Teaching–Learning-Based Optimization algorithm (TLBO-CWRNN-MN). Models were trained on pH, temperature, and total dissolved solids to enable rapid, cost-effective prediction of LSI without the need for complex laboratory analyses. While the LSTM model exhibited limited accuracy in early scenarios, its performance improved substantially with the inclusion of extended temporal inputs. Nonetheless, the hybrid architectures significantly outperformed the baseline models. In the most comprehensive scenario, TLBO-CWRNN-MN achieved the highest predictive performance (R = 0.982, RMSE = 0.054, NRMSE = 0.034), followed by CWRNN-MN (R = 0.972, RMSE = 0.079, NRMSE = 0.049), underscoring the effectiveness of integrating memory components and metaheuristic optimization in capturing nonlinear temporal dependencies. These findings are especially pertinent to regions such as West Azerbaijan Province, where escalating water scarcity and groundwater depletion have intensified ecological vulnerabilities, notably the decline of Lake Urmia. The proposed models offer a scalable and reliable approach to predictive water quality management, providing critical support for infrastructure sustainability and long-term environmental stewardship.