(92b) Comparison Between Data-Driven and Physics-Based Models of a Tabletop Analog to an Oil-Water Field

A 2009 study estimated that onshore US oil and gas fields alone produce 56 million barrels of water per day, including sources such as injected water used to increase production [1]. Most is reinjected locally or treated locally. The remainder is transported and disposed, typically through injection. Water remediation costs vary from as low as USD 0.07 to more than USD 4.00 per barrel, making water remediation a multibillion dollar per year industry [2]. Such high levels of water production greatly decrease the economic viability of oil fields.

Methods

A tabletop analog of an oilfield was designed to study techniques to mitigate water production. It represents a constrained section of an oilfield corresponding to a reservoir and producing well. A porous layer contains a mixture of oil and water situated between non-porous layers, and a miniature wellhead assembly extracts the oil-water mixture. The reservoir pressure is created by a hydrostatic source to produce flow from the well. The well includes sensors to measure the pressure, flow rate, and oil-water ratio of produced fluids.

The mitigation strategies require modeling that can predict production rates of oil and water. Both data-driven and physics-based models were developed to account for variations in reservoir pressure, oil-water ratio in the reservoir, production rates, and well-contact location relative to the oil-water interface. In this study, reservoir pressure, oil-water ratio, and wellhead pressure were varied and flow rates of oil and water measured.

Models

Two models are compared—one data driven and one physical. The input variables were varied through a series of 18 realizations over three reservoir pressures, three wellhead pressures, and two oil-water ratios. The production rates of the tabletop analog were recorded.

Physics-based Model:

The conditions of the well were simplified to facilitate calculations for one-dimensional flow of the oil-water mixture from the reservoir. Using conservation of mass and momentum equations, the production flow rates can be predicted and compared to the tabletop analog. Results were directly compared to the 18 realizations to verify the generality of the model.

Data-driven Model:

An artificial neural network was used to build the data model. It uses the supervised learning paradigm, whereby a training dataset is provided for the model to infer the mapping implied by the input and output variables in the dataset. The training dataset contains the three input variables: reservoir pressure, wellhead pressure, and the total oil and water volumes in the reservoir in the 18 realizations. The training dataset also contains the output variables—oil and water flow rates.

An additional validation dataset was used to tune the parameters of the data-driven model.

Finally, three test datasets, one with input parameters bounded by the training data and two with input parameters outside the training data, were used to verify if the model was generalizable by comparisons of predicted and measured production rates of oil and water.

Sensitivity

In practice, it is not possible to know reservoir composition with a high degree of certainty. Instead, estimations are calculated within some constrained error bound. To simulate this and measure its effect on the predictive nature of the data-driven model, input data are pre-processed to artificially introduce fluctuations in reservoir composition. The extents to which fluctuations in reservoir composition propagate through the two models are investigated.

References

[1] Clark, C.E. and Veil, J.A., Produced Water Volumes and Management Practices in the United States, Argonne National Laboratory report ANL/EVS/R-09/1 for the U.S. Department of Energy, 2009.

[2] Energy-Water Nexus: Information on the Quantity, Quality, and Management of Water Produced during Oil and Gas Production, U.S. Government Accountability Office, 2012.