Feature engineering improves cross-mode generalization in machine learning-based prediction of membrane capacitive deionization

In membrane capacitive deionization (MCDI), an energy-efficient alternative for low-salinity water treatment and desalination, the accurate prediction of output variables (e.g., effluent pH and ion concentration) is essential for process stability and optimization. Although machine learning and deep learning (DL) models have recently been applied to MCDI process data, their generalization performance is degraded under operating conditions that differ from training. To address this, we propose a feature engineering strategy that incorporates time-based statistical information into the input variables. Through sliding windows, statistical features were added to the input data, and a random forest (RF) model trained under constant-current (CC) operation was evaluated for predicting the effluent pH under constant-voltage (CV) conditions. The RF model with data features achieved an R 2 of 0.6…