Hybrid modelling for remote process monitoring and optimisation

Process simulation is used to develop a digital twin representation of chemical processes typically for process op-timization or what-if scenarios. However, it is known to be computationally expensive and there is an increasing need for process remote monitoring and optimization. This is where machine learning models shine, where they can run up to several orders of magnitude faster than their equivalent first principle process simulation models. Most of the previous work in this area tend to demonstrate the ability of machine learning models to accurately capture complex, non-linear relationships between process parameters in various chemical processes. Two impor-tant aspects are rarely discussed, namely the construction of machine learning models using operating data and the deployment of these models. In this paper, we address these two aspects and review the different informa-tion silos in industrial settings that need to be considered when working with hybrid models (combining process simulation and machine learning). This is illustrated by a case study for an operating natural gas dehydration unit, covering data management, process simulation, machine learning and visualization. We demonstrate how the hybrid models can be constructed and packaged as an online monitoring and a prediction dashboard. Several unique challenges are also highlighted including the reliability of field data and operational deviations due to operability and controllability -all of which need to be understood in order to successfully translate operating systems to process simulation and machine learning models that are reliable and accurate. While the exact imple-mentation may vary from project to project, the current work serves as an example and highlights the important considerations to make when working with such systems.