Multiphase flow metering provides an essential input to control and optimization systems for oil and gas production. Traditionally, flowrates have been measured by well tests which are associated with production losses, as they involve re-routing of the produced multiphase fluid streams to a test separator for estimation of single-phase flowrates. Alternatively, in-line multiphase flow meters have been applied. However, these sensors are costly to purchase and maintain in subsea operation. Another possible way to estimate multiphase flowrates is so called Virtual Flow Metering (VFM), where the physical flow meters are replaced by cheaper, more accurate sensors (such as pressure or temperature) giving input to a mathematical model of the process, that allows the flow rate to be calculated rather than measured. These new possibilities to model complex systems are also facilitated by the recent availability of computing power and advances in machine learning algorithms. The plan is to bring virtual flow metering to its full potential; that is, building more powerful process control and optimization systems with accurate and inexpensive flow measurements based on reliable sensors instead of using inaccurate and expensive physical flow meters.

Most of the models for VFM developed in the past can be broadly classified into two categories, i.e., first principles models and machine learning models. These two approaches have associated advantages and limitations. The first principles models require a detailed understanding of the underlying physical phenomena of the complex process system which may be difficult and time-consuming to develop, but offers higher trust in the model predictions. On the other hand, machine learning models do not require deep process understanding, but require larger amounts of input data and have poor generalizability, meaning their performance is limited to the characteristics of the specific set of training data used in their development. 

The goal of this project is to develop tools that combine machine learning methods with knowledge-based first principles models to develop hybrid models. These hybrid models will have higher predictive performance and can further be used to control and optimize overall production systems.