AI-driven MPC/DMC: Automated Model Generation & Smart Control for the Future Process Industry

Abstract

Traditionally, Model Predictive Control (MPC) controllers applied in refinery plants rely on linear prediction models derived from step response tests, under steady state operation. Modern process industries often produce large amounts of measurement data, which brings potential for AI-driven modeling instead. Through deep learning frameworks such as Long Short-Term Memory (LSTM), a model could learn system dynamics through historical input-output data. However, identifying dynamics can be challenging, since process data often is collected in closed-loop setting with an active controller.

This thesis investigates an LSTM-based MPC controller applied to a pass-balancing problem in an industrial furnace. The LSTM prediction model was trained using either real historical process data or synthetically generated open-loop data based on a First Order Plus Dead-Time (FOPDT) process model. A Hybrid Physics-Guided LSTM-MPC was also evaluated, utilizing the FOPDT model to ensure the correct physics of the system while applying the LSTM model for nonlinearities and disturbances.

The performance of the proposed LSTM-MPC controller was evaluated in comparison to the traditional linear MPC controller in a custom simulation environment. The results demonstrate that performance and stability of an AI-driven MPC strongly depend on the quality of the training data, particularly in closed-loop settings. The prediction performance in regards of MSE error of the LSTM network alone is not enough to determine if the model will do well in closed loop control, hence evaluation through simulation is essential. Furthermore, stable closed loop performance was achieved using synthetically generated training data, showing that LSTM-based MPC can be a promising approach for future nonlinear process control, and motivating further research into automated model generation for industrial control systems.