Towards Machine Learning of Power-2-Methanol Processes

Many dynamic models in process engineering rely on uncertain or even largely unknown mechanism. These parts are often modeled using best-practice knowledge or heuristics, which may result in 1) overly complex models, 2) models which do not reproduce the expected predictive outcome or 3) ill posed optimization problems. A promising approach is the use of machine learning surrogate models inside a partially known mechanistic model. In this study the impact of measurement noise and sample rate on the predictive performance of such a hybrid model is investigated. As an example, process methanol synthesis is used. Additionally, it is shown that the surrogate model can learn even unobservable states from the data.