A Novel Approach to Reservoir Simulation Using Supervised Learning

Abstract

Numerical reservoir simulation has been a fundamental tool in field development and planning. It has been used to replicate reservoir performance and study the effects of different field conditions in various reservoir management scenarios, and during field development and planning. Consequently, physics-based simulations have been heavily used during various reservoir studies such as history matching, uncertainty quantification and production optimisation; grid size and geological complexity also have a significant influence on the speed of the simulation. Furthermore, heterogeneities such as natural or hydraulic fractures can cause convergence problems and make the simulation even more time-consuming and computationally expensive. Due to being computationally demanding, such studies are also extremely time intensive. As a result of this downside, it is practically impossible to follow workflows such as the closed-loop reservoir management approach, which recommends updating the model every time a set of new data is available. Additionally, any management scenario must be approached from a business and economic standpoint. This means that, based on the predefined objectives within the study, the respective layers of precision must be chosen by the user. Therefore, if less expensive techniques can be implemented and provide adequate results, the use of more accurate and costly methods cannot be justified. One popular solution in overcoming this problem involves the creation of an approximate proxy model for the required features of the desired reservoir. This is achieved by either replacing or combining the physics-based model with this approximate model. However, by following this approach, the designed proxy model is only able to represent its corresponding reservoir, with a new proxy model needed to be rebuilt from scratch for any new reservoir. With consideration to the overall runtime, it can be observed that the time taken in iteratively running a numerical reservoir simulation may be faster than the time taken by the entire process of building, validating and using a proxy model. Therefore, this thesis focuses on the feasibility, advantages and contribution of a complete stand-alone AI-based simulator, Deep Net Simulator (DNS), in a wide range of different conventional and tight sand reservoir scenarios in 1D, 2D and 3D space. This thesis involves the use of deep learning to create a data-driven simulator, Deep Net Simulator (DNS), that enables the simulation of a wide range of reservoirs. Unlike conventional proxy approaches, a large amount of data is collected from multiple reservoirs with varying configurations and complexities. This results in the creation of a comprehensive database, including various possible reservoirs’ features and scenarios. The hypothesis is that such an approach will enable the data driven model to perceive and understand the principles that make up reservoir modelling and that the model will act as an excellent approximation to the equations that traditional physics-based numerical simulators solve. This objective is highly possible, since deep learning has been shown to be a great universal function estimator, which is capable of estimating the physics once given enough data and observations. Hence, this thesis aims to develop a series of data-driven models with the aforementioned features for various types of reservoirs. Initially, a workflow is designed to integrate a commercial simulator with a data extraction algorithm, enabling the generation of input-output simulation datasets. Next, the datasets are generated and reviewed. These datasets are then used in the training, validating and testing of the developed models. These developed data-driven models are able to learn and reproduce the physics governing fluid flow for a range of different scenarios: a single-phase oil reservoir in one-dimensional space, a single-phase gas reservoir in two-dimensional space, a single-phase gas reservoir in three-dimensional space, and hydraulically fractured tight gas reservoirs in two-dimensional space. The developed model was evaluated in terms of precision, speed, and reliability. For each scenario, the developed model was compared with a commercial reservoir simulator, and its performance was assessed using the following metrics: mean absolute error, mean absolute percentage error (MAPE), mean relative error, mean square error, root mean square error and r squared. The developed model was able to predict 45%, 70% and 90% of the cases with less than 5%, 10% and 15% MAPE, respectively. Furthermore, depending on the number of cells requiring outputs, the developed model was able to reduce runtime by 100% up to 1.04E+08%. This thesis takes the first steps towards establishing a new approach using AI and deep learning, for reservoir management procedure that is cheaper, less computationally demanding and more adaptable. This approach may result in a better value creation alongside a quicker decision-making process and, possibly, the advantage of integrating other attributes and data that are currently not used in physics-based models.