Neural Networks Tools for Improving Tacite Hydrodynamic Simulation of Multiphase Flow Behavior in Pipelines

Transient multiphase flow simulators are generally used to dimension the production scheme. One of the problems encountered is to predict accurately the pressure drop and the liquid holdup. This can be solved using an accurate numerical scheme and an appropriate thermodynamic behavior linked to an accurate and robust hydrodynamic model. In the Tacite Code, developed by IFP, a mechanistic hydrodynamic model has been developed. This model is able to predict the flow regime, the phase velocities and the local pressure drop for all slopes and all diameters. It contains closure laws based on flow regimes. This mechanistic model has been validated against various data banks. The two limitations of such an hydrodynamic model may be its mathematical disturbance (continuity, derivability are not always guaranteed) and the time consuming. This can be troublesome when using an accurate numerical scheme that requires derivative computation and for real time purposes. This paper presents a neural network based approach to efficiently replace the hydrodynamic module in the two phase model with the following two objectives: - to avoid discontinuity problems during hydrodynamic computations;- to reduce significantly computational time. This method was tested with experimental and simulated data. The results given in this paper prove the relevancy of this approach.