Machine Learning Applied to Multiphase Production Problems

In reservoir engineering, it is important to understand the behavior of the reservoir, often by seeing the dynamics of bottom hole pressure (pwf) and flow rate (q) in each well. Bottom hole pressure and flow rate are commonly modeled using physical model, for example, reservoir numerical simulator. This approach requires physical parameters, including rock properties and reservoir geometry, which we do not always know. In some cases, collecting those physical parameters and building the physical model might not be practical. Applying this approach in a field that has thousands of wells with a very dynamic environment requires a lot of effort.
Another approach uses historical bottom hole pressure and flow rate, which are commonly easier to obtain. Using machine learning algorithms, we can build a reservoir model by learning the historical pattern of bottom hole pressure and flow rate (or training set) without the physics of the flow being programmed explicitly. Using this model, we can predict flow rate given bottom hole pressure (or vice versa). We can also utilize this model as a diagnostic tool. For example, if liquid loading or condensate banking or wax/asphaltene deposition starts to happen, the actual pressure and flow rate response will be different than that suggested to the model, hence flagging that the reservoir performance has changed in character.
Previous efforts by Liu (2013) and Tian (2014), have been successful in applying machine learning method in single-phase cases. Important features were identified and gave satisfactory results in several different cases. The study found that although pressure and flow rate relationship can be nonlinear, the problem can be formulated as a linear problem and the nonlinearity is expressed in the features.
This research focused on solving multiwell and multiphase problems using machine/deep learning approaches. The study has identified important features in such problems. This study explored ten different machine/deep learning algorithms and their performance evaluation. Python is very helpful for this purpose as it has a lot of machine/deep learning libraries to choose from and is handier in solving machine learning related problems. The study also examined the impact of multiphase flow, reservoir heterogeneity, and data noise.