Scholarlink

<p>The complexity of influent water quality and the diversity of predictive models pose significant challenges to the efficient identification of optimal membrane flux prediction models. This study proposes an integrated decision tree approach combined with AutoML to classify influent types simultaneously (with an accuracy &gt; 96.08%) and rapidly screen the optimal model from 2400 candidates. By classifying the influent types under typical conditions such as groundwater, reservoir water, reclaimed water, and recycled water, the model optimization efficiencies increased by 43.59%, 52.35%, 51.52%, and 48.05%, respectively, compared with that of the unclassified treatment. The optimal model screened by AutoML demonstrated excellent predictive performance across varying influent parameter fluctuations (<i>R</i><sup>2</sup> &gt; 94.5%, RMSE &lt; 0.061), with a reduction in optimization iterations of up to 60.48%. Parameter importance analysis revealed that accurate matching between influent parameter variations and model characteristics is key to achieving high prediction accuracy. Furthermore, the selected optimal models enabled accurate prediction of the optimal scale of inhibitor dosage under different influent conditions, providing theoretical and technical support for the scientific application of membrane treatment chemicals.</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

Automated machine learning-based reverse osmosis membrane flux prediction and chemical dosage dynamic adjustment

  • 期刊论文

  • 发表日期：

    2026年1月5日

查看全文
  • Zhiyang Cheng,
  • Yang Yu,
  • Xia Meng,
  • Yating Wang,
  • Fangong Kong,
  • Minghua Zhou,
  • Jie Wang

摘要

The complexity of influent water quality and the diversity of predictive models pose significant challenges to the efficient identification of optimal membrane flux prediction models. This study proposes an integrated decision tree approach combined with AutoML to classify influent types simultaneously (with an accuracy > 96.08%) and rapidly screen the optimal model from 2400 candidates. By classifying the influent types under typical conditions such as groundwater, reservoir water, reclaimed water, and recycled water, the model optimization efficiencies increased by 43.59%, 52.35%, 51.52%, and 48.05%, respectively, compared with that of the unclassified treatment. The optimal model screened by AutoML demonstrated excellent predictive performance across varying influent parameter fluctuations (R2 > 94.5%, RMSE < 0.061), with a reduction in optimization iterations of up to 60.48%. Parameter importance analysis revealed that accurate matching between influent parameter variations and model characteristics is key to achieving high prediction accuracy. Furthermore, the selected optimal models enabled accurate prediction of the optimal scale of inhibitor dosage under different influent conditions, providing theoretical and technical support for the scientific application of membrane treatment chemicals.