(152g) A Deep Neural Net Model for Oil/Water Separation in Oil Production Pipelines

Authors: 
Tsai, K., Shell International E&P
Lu, L., Shell International E&P
High fidelity 3-D engineering simulations are valuable tools providing recommendations and solutions difficult to obtain otherwise. However, those simulations can be costly due to expensive computational requirements. In this study, a deep neural network model trained by computational fluid dynamics simulations capable of predicting oi/water separation in horizontal oil pipelines was developed to accelerate and replace the CFD simulations. The goal is to provide useful information for oil production operators to effectively determine the use of corrosion inhibitors, which can add significant cost to the operation, and incorrect decisions can have serious impact on our environment.

The CFD model was developed in the past 4 years for oil/water separation in horizontal pipelines to predict the water wetting probability on carbon steel pipeline surfaces [1]. The model has been shown to correctly reproduce the flow regime map widely used in industry across ranges of water cut, mixture velocity, pipe diameter, oil density, viscosity and surface tension. However, the model requires expert set up and can take days to complete. Recently deep learning neural network algorithms has gaining momentum in the way that it automates the correlation construction between input and output data and allows complex hyper-dimensional correlations to be built with high accuracy. The efficiency comes from the optimization procedure unique to convolutional neural network (CNN) models. By parametrizing the CFD simulations it is possible to pre-calculate the results in all the necessary parametric spaces and use them to train the CNN model.

The preliminary results in two parametric spaces, water cut and mixture velocity, are presented. The trained CNN model showed very high accuracy with the mean square errors lower than 0.01 in all cases. A total of ~4,000 data points were used to train the CNN model. The training time was around 4 minutes, and the inferencing time was less than 5 milliseconds, a performance gain of 106 in comparison to the CFD simulations.
The model is being extended to include the other parameters aforementioned. It can also be extended to model other types of predictions as well, such as erosion and corrosion, and with more complex geometries other than straight pipes. The possibilities are endless.

[1] K. Tsai, “Modeling the separation of oil and water in pipelines,” Paper 452a, 2017 AIChE Annual Meeting, Minneapolis, USA.

References

Hu, H. and Y.F. Cheng, “Modeling by computational fluid dynamics simulation of pipeline corrosion in CO2-containing oil-water two phase flow,” J. Petroleum Sci and Eng., 146, pp. 134-141, 2016.

Kanwar, S., Study and modeling of sweet corrosion of multiphase mixtures in horizontal pipelines, Master of Science Thesis, Ohio University, USA, 1994.

Tsai, K., “Corrosion modeling using electrochemistry and computational fluid dynamics,” Paper No. 355b, 2017 AIChE Annual Meeting, Minneapolis, MN, USA., Oct., 2017a

Tsai, K,, “Modeling the separation of oil and water in pipelines,” Paper No. 452a, 2017 AIChE Annual Meeting, Minneapolis, MN, USA, Oct. 2017b

Zhang, G.A. and Y. F. Cheng, “Electrochemical characterization and computational fluid dynamics simulation of flow-accelerated corrosion of X65 steel in a CO2-saturated oilfield formation water,” Corros. Sci. 52, pp. 2716-2724, 2010.

Checkout

This paper has an Extended Abstract file available; you must purchase the conference proceedings to access it.

Checkout

Do you already own this?

Pricing


Individuals

AIChE Members $150.00
AIChE Graduate Student Members Free
AIChE Undergraduate Student Members Free
Non-Members $225.00