(164e) Enhancing Extrapolation Capabilities of a Data-Driven Prediction Model for an Air Separation Unit Using a Digital Twin

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.

However, data-driven models generally show weaknesses in their extrapolation capabilities, i.e., in areas where no data was available for model training. In this work, a digital twin is therefore deployed to act as an additional data source. By using the digital twin, data-poor operating areas of the plant can be augmented with artificially generated data. The prediction model is then trained with data from both data sources, i.e., historical plant data and data produced with the digital twin. One requirement is, though, that the digital twin must be modeled in close accordance to the plant. Another important issue to address is the plant-model mismatch occurring due to measurement errors at the plant.

The aim of this work is to determine the prediction quality of a hybrid prediction model, trained from different data sources, and compare it with a fully data-driven prediction model. Especially in the data-poor areas, it is expected to achieve a control model with an increased prediction quality. In addition, the uncertainty behavior of the control model is investigated to determine whether this is improved by augmenting the original plant data set.

The hybrid model can be particularly advantageous for newly constructed plants as the time required to acquire a meaningful data set for training the data-driven prediction model can be drastically reduced. This would allow the plant to be operated at high performance more quickly.