Comparative Performance Evaluation of Machine Learning Models for Predicting Water Yields in Desiccant-Driven Water Harvesting System

Global warming, climate change, and geopolitical conflicts have collectively intensified the global challenge of water scarcity, making access to clean water a critical issue that demands urgent attention. To address this crisis and ensure equitable access to water for all, innovative solutions and advanced technologies are necessary. In response to this growing concern, this study explores the potential of Desiccant-driven Water Production Systems as vital components in sustainable resource and energy management strategies. Specifically, we evaluate the use of cutting-edge machine learning models for Atmospheric Water Harvesting (AWH) systems, focusing on predicting water production efficiency across four key targets: Water Production (WP), Cumulative Water Production (CWP), Log-Transformed Water Production (WP-log), and Log-Transformed Cumulative Water Production (CWP-log). Through the application of advanced machine learning techniques, including DNN, K-NN, XGBoost, LGBM, AdaBoost, SVM, and RF, we assess their predictive performance based on three essential metrics: R² score, Mean Absolute Error (MAE), and Mean Squared Error (MSE). Our results demonstrate the significant benefits of log transformations in improving model accuracy, particularly for predicting CWP, which is influenced by cumulative uncertainties and fluctuations over time. DNN model achieved impressive R² scores of 0.95 for WP-log and 0.97 for CWP-log, while SVM achieved even higher R² scores of 0.96 for both log-transformed targets, highlighting its ability to capture the complex non-linear relationships that govern AWH systems. In terms of error reduction, SVM led with the lowest MAE of 0.028 and MSE of 0.0023 for both log-transformed targets, showcasing the model's potential to minimize prediction errors, even in highly variable environments. The application of these advanced machine learning models in AWH systems is crucial for optimizing water production predictions, particularly in environments that experience significant fluctuations. Our findings demonstrate that machine learning, combined with preprocessing techniques like log transformation, can significantly enhance the accuracy of water production forecasting. This capability is essential for improving the operational efficiency and sustainability of AWH systems. These results underline the transformative potential of machine learning in AWH applications and provide valuable insights into the future development and optimization of water harvesting technologies to address the global water scarcity challenge.