Implementing a Hybrid Digital Twin Approach in a Crude Unit Operation

Abstract

Abstract Over the past decade, Machine Learning or, more generally, Artificial Intelligence, has made a stelar entry into the O&G world and is today routinely used from exploration all the way up to retail. At the same time, physics-based tools are still being seen as key for value generation and the O&G community is increasingly looking at combining "traditional" with "new" technologies in what has come to be known as "hybrid" tools. In this paper we explain how we Baker Hughes, C3 and KBC are combining Physics and Machine Learning to create Hybrid Digital Twins that help operators improve their margins, reduce their emissions and, in general, position themselves for sustainable growth. Before they are in operation, plants (and oilfields alike) are typically simulated using physics-based tools. In practice, therefore, physics-based simulation models pre-date any plant data. Once plant data starts being produced, however, it becomes clear they do not match exactly the theoretical models that were used before operation. It also becomes clear that plants are operated in a very narrow range, due to quality, production and HSE requirements. In this paper we document a hybrid approach for a digital twin used to generate optimized targets to a Crude Unit operation: synthetic data has been generated to overcome the limitations of available plant data. These synthetic data have then been used as additional training input for ML models, complementing instrumentation data, while at the same time, instrumentation data is used for calibration of the physics-based model. This hybrid approach has been applied to a Crude Unit, in an optimization use case, with remarkable results. The final goal was to make energy optimization targets to operations through its machine learning algorithms which are based on two years of historical data optimized by a rigorous process simulator. The targets are manually downloaded to the unit multivariable controller by the operators. The hybrid approach has been found to combine the benefits of each type of technique: First Principles provides a sound bases for (limited) extrapolation, while Machine Learning ensures the First Principles models remain tuned to the reality of the process and complements the physics in areas where it is insufficient to model the reality (as is the case of equipment performance degradation/failure prediction). The range of what-if studies (and, correspondingly, optimization options) is thus radically extended. This paper shows incremental benefits in optimization cases illustrated by a Crude Unit use case, resulting from the application of a hybrid digital twin using both physics-based and machine-learning models. The hybrid approach allows optimizers to evaluate scenarios beyond the historical operating envelope and opens the door to incorporating equipment condition considerations.