Systematic analysis of membrane capacitive deionization using stepwise operating data and multiple machine learning models: The effect of data attributes on process estimation

Understanding the characteristics of operational steps in membrane capacitive deionization (MCDI) is essential for effective process optimization. This study analyzed MCDI operating data collected from five distinct operational steps, applying six optimized system adjustment variables that achieved a balance between energy consumption and water productivity. As a result, stable MCDI operation was confirmed, with a specific energy consumption (SEC) of approximately 1.82 kWh/m³ and a permeate flow rate of 12.5 L/min. The MCDI data collected during system variable manipulation were used for a detailed analysis of the impacts of operational steps on process performance estimation. Particularly, this study focused on evaluating the role of categorical MCDI data that represent operational steps—an aspect that had not been explored previously. Seventeen machine learning (ML) models were employed and tested using three data types: continuous, categorical, and complete (combining both types). The results indicated that incorporating categorical data, initially absent in the raw dataset, significantly enhanced predictive accuracy compared to continuous data alone. Furthermore, the ML models exhibited higher accuracy in estimating outflow total dissolved solids (TDS) than electrical energy (EE), due to the limited variation of EE across operational steps. Additionally, this study evaluated the efficacy of hybrid ML approaches using categorical data, which revealed that voting regression outperformed stacking regression for predicting MCDI operating data. These findings highlight the significance of meticulous and stepwise operational analyses and underscore the benefits of incorporating categorical data for improved MCDI process management and optimization.