The research findings have been published in the article Predictive Model for Bottomhole Pressure Based on Machine Learning.
Machine learning is increasingly used in various industries. However, the oil industry is characterized by the high degree of unpredictability of its processes. Therefore, there were initially concerns that it would not be possible to apply machine learning in this case. Nevertheless, researchers from HSE and Schlumberger have managed to create and train an artificial neural network to predict bottomhole pressure. For the first time ever, the neural network has been successfully used to analyze essentially non-stationary processes which are specific to multiphase flows in wells.
Bottomhole pressure can change considerably over time, especially in the early stages of putting the wells into operation. This is due to multiphase flows, when water, oil and gas flow through the wellbore concurrently, in changing proportions and at various velocities. Under certain conditions, the slug flow regime can be developed in the wells, when liquid and gas are delivered to the surface successively as plugs, thus causing considerable pressure fluctuations throughout the system. ‘Knowledge of the pressure changing over time and bottomhole pressure management are critical to ensuring that the well is put on stream in a safe and optimal way. For example, to ensure that an efficient well clean-up strategy is chosen during the initial well start-up,’ explains Pavel Spesivtsev, Project Manager at the Schlumberger Moscow Research Center. Bottomhole pressure prediction also helps to efficiently manage the inflow at the well bottom in order to increase oil recovery.
The proposed machine learning algorithm imitates the presence of an experienced expert who can predict the behavior of the hydrocarbon (HC) mixture at the well bottom by observing the HC and water at the surface.
A specially developed physical model was used for calculations with parameters corresponding to processes which occur in oil producing wells.
‘The objective was to gain an understanding of what goes on at the bottomhole by looking at the processes at surface’, explains Alexey Umnov, a researcher from HSE Faculty of Computer Science. ‘To achieve this, we have developed almost 3.5 thousand artificial well behavior scenarios’.
The scenarios took into account not only the behavior of a single well, but also a large data array, obtained from a number of wells which were operating in various regimes and with different control parameters. A prediction was then yielded for a particular parameter – bottomhole pressure.
After training, the neural network helps to predict bottomhole pressure with a minimal computational burden. This is an advantage compared to conventional approaches based on numerical simulation. The researchers showed that the bottomhole pressure can even be predicted for substantially non-stationary slug flow of liquid and gas. For most predictions, the mean square error was less than 5%, which is an excellent quantitative result.