(230e) Enhancing Air Separation Unit Control: A Multi-Model Approach for MPC with Argon Production

In air separation units, improvement of plant control, e.g., with the usage of model predictive control (MPC), can lead to an increase in plant efficiency. However, the physical modeling of non-linear prediction models used here is often time-consuming and too complex to be used on-line in real-time operation. Instead of physical modeling, though, non-linear prediction models can also be modeled using a fully data-driven approach. An artificial neural network (ANN) used here only requires historical plant data and can therefore learn the non-linear plant behavior itself. In this work, the fully data-driven approach used for the control of an air separation unit will be presented while the higher-level MPC controls various control loops.

In addition to the pure products nitrogen and oxygen, the air separation unit under consideration also produces argon. Additional columns and plant components are required for this and are taken into account in the process topology. Argon production is now also to be controlled by the MPC, thus including a further control loop. Unfortunately, it is in the nature of the process that the system with argon separation has very high response times. This means that changes in the process take much time to become visible in the process and relevant in the production of the product. For this reason, the argon control loop cannot simply be built into the existing data-driven control system, as the current prediction horizon is not sufficient for this. It is also to be mentioned that the used prediction horizon needs to be equal for all installed control loops.

Although the prediction horizon could now be increased for all control loops, this would significantly increase the overall complexity of the model, resulting in huge model training costs. Instead, a new multi-model approach is to be developed, whereby two separate prediction models are used in parallel. While the first is identical to the previous setup and uses the existing control loops, the second prediction model will only include the argon system. This approach allows the size of the prediction horizon to be flexibly adapted for the respective subsystem. The aim is to show that with this multi-model approach, the MPC used is now also able to control the argon system and thus achieve increased system efficiency compared to the previous controller setup.