Scholarlink

<p>The global water crisis continues to intensify due to climate change, rapid population growth, and industrialization, necessitating innovative solutions for sustainable water production. Atmospheric Water Harvesting (AWH) has emerged as a promising method to extract water from ambient air, leveraging desiccant-based adsorption and condensation processes. However, optimizing AWH efficiency requires accurate predictive models that account for complex environmental and operational factors. This study explores the application of Deep Neural Networks (DNNs) to predict Water Production (WP) and Cumulative Water Production (CWP) in an AWH system utilizing calcium chloride (CaCl₂) as the primary desiccant. A comprehensive dataset encompassing humidity, desiccant concentration, machine slope, glass temperature, bed temperature, and other key variables was used to train and evaluate multiple DNN architectures. Extensive data preprocessing techniques were employed, including outlier detection, feature normalization, and logarithmic transformation to improve model generalization. Our results demonstrated that log-transformed target variables significantly enhanced predictive performance, with DNN (WP-log) achieving an R² score of 0.95 and DNN (CWP-log) attaining an R² of 0.97, compared to lower performance in non-transformed models. Additionally, Mean Absolute Error (MAE) and Mean Squared Error (MSE) scores confirmed that log-transformed models reduced prediction errors, achieving MAE values of 0.032 and 0.028 for WP-log and CWP-log, respectively. These findings highlight the importance of nonlinear modeling approaches and data transformation in achieving reliable AWH predictions. This study represents a novel contribution to the field of AI-driven water harvesting by demonstrating the superior predictive capability of DNNs over conventional regression models. The results establish a data-driven optimization framework that can inform real-time adjustments in AWH systems, maximizing water collection efficiency under varying environmental conditions. The integration of machine learning in AWH technologies marks a significant advancement toward scalable, sustainable, and intelligent water harvesting systems. By bridging AI-driven predictive analytics with environmental sustainability, this study lays the groundwork for the next generation of adaptive water harvesting technologies, contributing to the global effort to address water scarcity through smart, efficient, and cost-effective solutions.</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

Fine-Tunning Deep Neural Networks for Performance Optimization in Desiccant-Assisted Atmospheric Water Harvesting

  • 期刊论文

  • 发表日期：

    2026年1月19日

查看全文
  • Mehrdad Ghiyasvand,
  • Elshan Soltani

摘要

The global water crisis continues to intensify due to climate change, rapid population growth, and industrialization, necessitating innovative solutions for sustainable water production. Atmospheric Water Harvesting (AWH) has emerged as a promising method to extract water from ambient air, leveraging desiccant-based adsorption and condensation processes. However, optimizing AWH efficiency requires accurate predictive models that account for complex environmental and operational factors. This study explores the application of Deep Neural Networks (DNNs) to predict Water Production (WP) and Cumulative Water Production (CWP) in an AWH system utilizing calcium chloride (CaCl₂) as the primary desiccant. A comprehensive dataset encompassing humidity, desiccant concentration, machine slope, glass temperature, bed temperature, and other key variables was used to train and evaluate multiple DNN architectures. Extensive data preprocessing techniques were employed, including outlier detection, feature normalization, and logarithmic transformation to improve model generalization. Our results demonstrated that log-transformed target variables significantly enhanced predictive performance, with DNN (WP-log) achieving an R² score of 0.95 and DNN (CWP-log) attaining an R² of 0.97, compared to lower performance in non-transformed models. Additionally, Mean Absolute Error (MAE) and Mean Squared Error (MSE) scores confirmed that log-transformed models reduced prediction errors, achieving MAE values of 0.032 and 0.028 for WP-log and CWP-log, respectively. These findings highlight the importance of nonlinear modeling approaches and data transformation in achieving reliable AWH predictions. This study represents a novel contribution to the field of AI-driven water harvesting by demonstrating the superior predictive capability of DNNs over conventional regression models. The results establish a data-driven optimization framework that can inform real-time adjustments in AWH systems, maximizing water collection efficiency under varying environmental conditions. The integration of machine learning in AWH technologies marks a significant advancement toward scalable, sustainable, and intelligent water harvesting systems. By bridging AI-driven predictive analytics with environmental sustainability, this study lays the groundwork for the next generation of adaptive water harvesting technologies, contributing to the global effort to address water scarcity through smart, efficient, and cost-effective solutions.