Control of extractive dividing wall column using model predictive control based on long short-term memory networks

As a popular process intensification technology, extractive dividing wall column (EDWC) exhibits highly nonlinear behavior and poor controllability. Model predictive control (MPC) is recognized as an effective advanced control framework for managing EDWC, and the accuracy of the prediction model is crucial for the performance of the MPC controller. This work proposes a systematic framework for implementing long shortterm memory (LSTM) networks-based MPC to enhance the control performance of EDWC with the characteristic of multi-input multi-outputs, nonlinearities, and time delays. The case study models an EDWC for separating a toluene (TL) and 2-methoxyethanol (2-ME) azetropic mixture, which is optimized using the multi-objective particle swarm optimization (MOPSO) algorithm to decide optimal steady-state points for the control objective of LSTM-MPC. Subsequently, three different temperature inferential control (TIC) schemes are compared to select good input features for LSTM-MPC applications in this high-dimensional system. Extensive time-series data are collected to train an LSTM model with minimal mean squared error (MSE), and the final predicted trajectory demonstrates that the LSTM has a good generalization ability to capture the dynamic features of nonlinear EDWC. Two industrial disturbances are introduced to test the controllability of the LSTM-based MPC for EDWC. The simulation results show that the LSTM-MPC has better closed-loop controllability, with small offsets, negligible oscillation, and short transition time compared to the TIC schemes.