Application of kNN-GRNN algorithm for liquid holdup determination in two-phase flow

In some industries, such as chemical industry, nuclear power plants, geothermal energy, petroleum industry, etc. two-phase flow dynamics arise. Especially in oil and gas industry while producing hydrocarbons (oil and gas), starting with production string in the well through pipelines in the surface two-phase flow is transpired. Two phase flow is defined as two distinct phases are flowing simultaneously in the pipe. The dynamics affects the interphase between involved phases thus influences the distribution of both phases through the flowing pipeline. Relative distribution between phases is defined as holdup, generally defined as liquid phase holdup, or more precisely liquid holdup. This holdup is important for flow dynamics. The estimation of liquid holdup in pipelines is very substantial in the oil and gas industry. It affects seperation design equipment and slug catching. For both horizontal, inclined and vertical conditions various models are proposed to find this term in flowing conditions. Most of the proposed studies are for specific conditions and could not be generalized to other conditions. The start of the procedure in those models are identifying the flow pattern first (each flow pattern have its own solution procedure for finding liquid holdup) then finding the liquid holdup for each case. Liquid holdup is used in friction factor calculation and as a result for finding the pressure distribution along the pipeline. Flow pattern dependent models have the difficulty for finding the accurate liquid holdup value across the flow pattern transition boundaries. Those blurred transition region phenomena may result inconsistency as flow conditions differ. Neural Networks are designed by the inspiration of neurons in living organism nervous system. Neural networks are processors for understanding the data and diagnose the relations within them thus consequently find out and solve chaotic patterns related to data. Those systems foster themselves by getting every new data in time and do not need specific problem related programming. Neural networks evolve pertinent specifications and use them as the solution process. In recent years, they are used as solution process in many industries and in many engineering problems. Compared to conventional methods (empirical or mechanistic methods) Neural networks achieve promising results in various engineering applications such as multiphase flow modeling. Flow pattern identification, liquid holdup calculation and pressure distribution calculation can be conducted in multiphase flow modeling. In this study, machine learning algorithms were applied for liquid holdup determination. We tried to find the most important parameters that affect the two phase flow by trying all combinations of parameters used in the experiments. Mathematical correlations and Neural network algorithms were applied to the problem to compare K-nearest neighbours and GRNN combined (kNN-GRNN) algorithm. K-nearest neighbours was used as clustering for GRNN algorithm. Coefficient of determination (R) and average absolute percent relative error (AAPE) are used to compare the performance of all models. Hybrid kNN-GRNN algorithm performed the best with a 0.9573 R value and 18.02% AAPE.