Efficient process optimisation using AI-powered advanced process control

How Neural ODEs overcome the limitations of classical MPC and unlock new potential for process automation

Abstract

This paper demonstrates how neural ordinary differential equations (neural ODEs) can improve model predictive control in the process industry. The hybrid approach presented here combines the data efficiency of linear models with the ability to capture nonlinear dynamics. Neural ODEs model processes continuously, incorporate prior physical knowledge, and provide robust predictions even with limited data. Using a
nonlinear three-tank system as an example, it is shown that a neural ODE-based MPC achieves higher prediction and control performance than a linear MPC when using the same dataset.