Data-Driven Nonlinear Model Predictive Control for Air Separation Units

Abstract

In this work, a nonlinear model predictive control framework using a data-driven prediction model is used to control an air separation unit (ASU). The framework is integrated into an industrial automation platform, providing real-time control irrespective of the base layer control system vendor. The approach is a scalable solution that requires minimal effort in model training, leveraging an automated training framework. It is a powerful alternative to physically modeled nonlinear model predictive control (NMPC) and has already been validated by controlling several in-production-critical Linde ASUs. Additionally, a multimodel approach is introduced that combines several data-driven prediction models to only exploit the main dependencies of the considered process variables. Furthermore, the approach enables fast implementation at the plant as step tests are not required, which are usually time-consuming and interfere with the normal operation of the plant. Based on two applications, the control performance is presented for different load change scenarios, ensuring compliance with all product specifications. Furthermore, different operating modes, including different products, different equipment, and a different overall control task, can be captured by a single nonlinear model.