Combining Dimensional Analysis and Neural Networks to Improve Flow Assurance Simulations

When developing mathematical models for two- and three-phase flow in long pipelines, the most difficult challenge is to model the frictions, volume fractions and flow regimes accurately. This paper combines 3 different methods when confronting that problem: Dimensional analysis, mechanistic models, and Neural Networks (NN). It is shown that those methods supplement each other in important ways. The dimensional analysis is helpful in upscaling laboratory measurements to full-scale flowlines. In case of 3-phase gas oil water flow, the number of dimensionless groups turn out to be 14. NNs offer a way of correlating that many variables, and that allows the model to account for all dimensionless groups for all types of flow. That overcomes the limitations inherent in the more common practice of focusing on only a few parameters or dimensionless groups for each type of flow regime. But introducing NNs creates a new challenge: We need data to train them. The last problem is partly dealt with by building on well-established mechanistic models with various factors inserted. It is those factors which are trained by the NNs, not the dependent dimensionless groups themselves. Using mechanistic models rather than a pure "black box" approach leads to much faster training and more accurate results. That has made it possible to train the NNs based on a more moderate and therefore realistic amount of data than would otherwise be required. The novel approach has been used to develop new software. The FlowRegimeEngine, as it is called, is now incorporated in several steady-state and transient commercially available computer codes. At the end of this presentation results from one of them, FlowlinePro, have been compared to results from the well-established computer code OLGA. The results turned out to be very similar. The presented dimensional analysis also provides an interesting way of testing commercial software by checking whether results are dimensionally consistent. When doing steady-state simulations with two or more different data sets, chosen so that they form the same independent non-dimensional groups, the resulting dependent dimensionless groups should come out identical. If they do not, it is reason to treat the results with suspicion. When applying the test to FlowlinePro and OLGA, they both passed it nicely for the data-sets chosen here.