Comparison of Methods for Stochastic Multiphase Flow Rate Estimation

Abstract

Abstract Even though flow rates of oil, gas and water are the most important parameters of any hydrocarbon exploitation project, only a minor number of wells are permanently connected to a multiphase flow meter (MFM). Production well testing through a test separator is by far the most common practice to measure flow rates. This method comes with the restriction that only one well at a time can be tested, leading to sparse measurements over the production history of a well. In addition, operating conditions of the test separator can differ from flow line conditions and results in a discrepancy of flow line rates versus test rates. To account for this effect the well-test results have to be reconciled, usually backwards from the sales/export meter to each well. More and more wells are equipped with SCADA systems, measuring wellhead parameters in real-time. Data driven technologies are able to learn the correlation between well test flow rates and wellhead measurements, so that they can calculate production rates at any time between well tests. Data driven methods are fast, so that they can be applied in real time and are therefore sometimes called virtual meters. They can overcome the difference of well test and flow line conditions and predict fairly accurate rates. This paper compares a very simple regression method (MLR), a classification method (random forest) and a back-propagation neural network (ANN). All three methods are trained on real-time data from step rate well tests obtained from three wells. A blind test is performed on all three trained models to compare the predictability of each method. As the calculation time of these data driven methods is very short, they have been embedded in a stochastic loop to account for the variation of the acquired real time parameters during the tests. A standard deviation is calculated from the stochastic simulation. The standard deviation is not only a good measure of the predicting capability of each method, but also embraces all operational variations of flow conditions during testing. This leads to a better estimation of the most likely flow rate at current conditions.