Nonlinear Autoregressive Network with Exogenous Inputs for Dynamic Modeling and Control of Electrochemical CO2 Regeneration Cells

• NARX model captures nonlinear dynamics of electrochemical CO 2 regeneration cells • Partial Autocorrelation guides optimal NARX architecture selection • Kalman filtering of experimental data enhances data quality for model tuning • Tuned PI controller implemented for set point tracking and disturbance rejection Electrochemical CO 2 regeneration is a promising alternative to conventional thermal processes, offering lower energy requirements and compatibility with renewable electricity. However, the dynamic behavior of electrochemical cells is complex, nonlinear, and not yet well understood, making robust control design challenging. This study proposes a multi-step, data-driven methodology for dynamic modeling and control of an electrochemical CO 2 regeneration cell, using experimental data from two cathode spacer geometries. Raw experimental signals were preprocessed using a Kalman filter to suppress high-frequency noise. Partial autocorrelation analysis guided the selection of autoregressive orders, and a systematic grid search identified optimal nonlinear autoregressive models with exogenous inputs (NARX). A comprehensive methodology is presented that connects data preprocessing, model architecture selection and controller optimization into a coherent pipeline. The best-performing models achieved a greater than 97% fit and root mean squared (RMSE) values below 0.005 for both geometries, accurately reproducing transient CO 2 flux responses to stepwise current excitation. The analysis revealed that the rhomboidal spacer cell configuration exhibits longer memory effects and allows more consistent system identification compared to the parallel spacer. Closed-loop control studies with a discrete-time PI controller, tuned via a genetic algorithm, demonstrated accurate set-point tracking, disturbance rejection, and safe operation under actuator saturation. These findings highlight that NARX-based system identification not only enables accurate control-oriented modeling but also provides new insights into how cell geometry influences dynamic behavior.