Machine learning-based predictive control using on-line model linearization: Application to an experimental electrochemical reactor

Abstract

The electrochemical reaction-based process, a new type of chemical process that can generate valuable products using renewable electricity, is a sustainable alternative to the traditional chemical manufacturing processes. One promising research area of electrochemical reaction processing is to reduce carbon dioxide (CO2) into carbon-based products, which can contribute to closing the carbon cycle if CO2 is directly captured from the atmosphere. In this work, we demonstrate a model predictive control (MPC) scheme that uses a neural network (NN) model as the process model to implement real-time multi-input-multi-output (MIMO) control in an electrochemical reactor for CO2 reduction. Specifically, a long short-term memory network (LSTM) model is developed using historical experimental data of the electrochemical reactor to capture the nonlinear input-output relationship as an alternative to the complex, first principles-based model. Furthermore, the Koopman operator method is used to linearize the LSTM model to reduce the nonlinear optimization step in the MPC to a well-understood and easy-to-solve quadratic programming (QP) problem. The performance of the LSTM model, Koopman-based optimization, and MPC using the linearization of the LSTM model are evaluated with various simulations as well as open-loop and closed-loop experiments. As the results, the proposed MPC scheme can drive the two output states, that are concentrations of the products (C2H4 and CO), to their desired setpoints by computing optimal input variables (surface potential and electrode rotation speed) in real-time in closed-loop experiments. Furthermore, a transfer learning-based method is utilized to update the NN model to handle process variability.