Flexible risers might be subject to large pressure and density fluctuations when transporting multiphase mixtures as different flow pattern may take place. Those fluctuations might induce large forces on the structure which in turn may modify the global geometry of the riser. In this paper, results from one-way and two-way coupled fluid-structure simulations are compared against small scale experiments of a flexible lazy wave riser transporting gas-liquid flow. Different gas flow rates are tested in order to obtain a range of slug frequencies. Predicted values of pipeline pressure, top tension and riser displacement are compared with experimental data. The numerical prediction of the coupled behavior shows fairly good agreement with the experiments, having bias from 5 to 25% in slug frequency.

No abstract

Conventional plunger lifting is a transient process that consists of cyclic openings and closings of a gas well. Because of this complex behavior, using traditional physics-based models to simulate the coupled behavior of reservoir and wellbore performance is computationally rigorous and challenging. Therefore, this study proposes a machine learning-based approach to simulate gas production from plunger-lifted wells and facilitate the optimization of this process. The model is developed and validated using field data. Typically, high-frequency (1 minute) measurements of plunger arrival time as well as casing and tubing pressure, are available in a plunger-lifted well. In addition, some wells are equipped with individual high-frequency measurements of gas flow rate. However, in most cases, there is a single gas flow rate meter available for the entire well pad. Therefore, a machine learning methodology is formulated with input variables that include plunger arrival time, tubing and casing pressure, and instantaneous gas flow rate as an output variable. Due to practical considerations regarding plunger lift operation, the approach assumes that a training set (one week) is smaller than a testing set (one month). A feed-forward neural network model is trained and is found to provide results with acceptable accuracy. The architecture of the network is obtained by performing a grid search and by minimizing a mean squared error. In the next step, obtained gas production is treated as a function of "on" (opening) and "off" (closing) time periods. The objective of the second model is to reproduce the data and to construct a response surface by varying "on" and "off" time periods. Based on the results from several plunger-lifted gas wells, both models have a unified architecture that requires tuning weight coefficients with a training/development dataset. The neural network model to simulate the gas flow rate performs well; it is evaluated with common statistical parameters. The model requires gas flow rate measurements from routine production tests to build the training set. Having a gas flow rate model provides the opportunity to train another machine learning model as a function of "on" and "off" time periods. The new model is validated using the data during the final week of production history. The relative error between the data and the model is approximately 10%, which ensures the reliability of the model. A surface response is constructed over a range of "on" and "off" time periods to find an optimum point maximizing total gas production during the validation period (final week). Optimization results demonstrate that "off" time (fall + buildup) should be minimized, and "on" time (upstroke + after-flow) should be at a certain threshold. Current industry practice to optimize plunger lift cycles is based on factors such as average plunger rise velocity, and load factor. However, these methods do not optimize the actual variable of interest that is gas production. The unique contribution of the proposed approach is that it provides a robust tool to monitor the gas flow rate from an individual plunger-lifted well (flow rate allocation) and to optimize plunger lift cycles based on cumulative gas production. The model runs fast and can complement existing alarm systems on SCADA to adjust controller set-points in real-time.

A Gulf of Mexico oil field has six subsea producing wells that require production allocation. The wells are not equipped with multiphase flow meters and therefore, an indication of production through a virtual flow meter (VFM) is desired. The complexity is that in this field, well tests are performed as infrequently as semi-annually. The accuracy of the developed VFM is evaluated using well test data held out for validation. This paper talks about a near real-time cloud-based VFM solution. The approach is based on combining well measurements with fundamental flow equations through the tubing and choke. Well test data is used to calibrate these equations by tuning parameters such as the friction factor and choke flow coefficient. The models are auto-tuned based on the incoming and past well test data. Additionally, the models require a fluid properties table and wellbore deviation survey. When performing predictions, typical real-time measurements such as downhole and well head pressure/temperature and choke position are utilized. The approach was tested on a well in a nearby field which had more frequent well tests and then extended to the considered field, which had very few calibration data points (~2 well tests/year). The results were validated using left-out well tests and showed mean absolute percentage (MAPE) errors within 10%. Current industry practice for VFMs is to tune the model using the latest well test and use correction factors to calibrate frictional and gravity terms, which leads to predictions biased towards the latest well test. In addition to addressing these issues, this paper also outlines the auto-tuned near real-time implementation of the system. The utilization of hydrodynamic models as a basis also allow us to extrapolate beyond the range of the calibration dataset with good accuracy.