| dc.description.abstract | Simulation of reservoir models is a tool to optimize the development of an oil and gas reservoir. Part of the development is placement of wells in the reservoir, and this well placement optimization process is performed with different levels of automation, from manual search to algorithm-based optimization. To date, the most popular method of describing a well has been as a straight line defined by six decision variables (either in hybrid-spherical or Cartesian coordinates). This way of describing a well is beneficial because it has a compact formulation with a low number of decision variables that limit the degree of freedom in the optimization problem. Although this simple linear well representation will suffice in many cases, a straight line may be unsuitable for representing a realistic well trajectory if the topography is complex and the reservoir properties are heterogeneous. In this dissertation and associated articles, we present a methodology that increases the flexibility of the well description by dynamically changing the well path based on heterogeneous flow properties and topography. This new method is based on machine learning for automatic design of well paths in reservoir models. The method has been named Automatic Well Planner (AWP).
The methodology is based on innovations in machine learning to convey information from a simulated environment to a decision maker. The goal of the procedure is to create a realistic well trajectory and increase productivity compared to less flexible well descriptions. We created a computationally efficient simulation environment for a virtual drilling procedure. In this virtual drilling procedure, we use reinforcing learning to train a neural network (the decision maker) in how to design a well path that meets the predetermined specifications. Inspired by modern geosteering operations, the simulated drilling procedure develops the well trajectory sequentially. The decision maker computes a decision for each iteration on how to adjust the well trajectory based on information gathered within a distance-limited information horizon.
In this dissertation we will go through each individual component of the well planning algorithm, we present tests of the methodology using several visual, statistical, and dynamic tests. Combined the tests show that the methodology can learn to steer the well path according to the local reservoir properties, and it is able to develop new well trajectories with a low computational cost.
A potential application for this methodology in field development optimization. To investigate this further, we integrated the AWP into an optimization framework, where AWP acts as a drop-in replacement for the existing straight line well, using the same six decision variables as given by the endpoints for the straight line well. Replacing the well description was tested through increasing complexity, from a single producer in a deterministic reservoir model to several producers with an ensemble of reservoir models. Furthermore, we integrated AWP into a robust optimization procedure where we design wells based on geological information from the individual realizations in an ensemble. These tests show that in most cases, AWP enables a more efficient search and helps the optimization algorithm to earlier convergence and higher objective function values compared to the linear well description.
The methodology is also applied to reduce the number of decision variables for complex well trajectories. We retain the main aspects of the introduced methodology; the ability of the algorithm to adapt to the different geological circumstances. Additionally, we increase the flexibility of this version of the AWP to allow the methodology to make larger changes to the well trajectory design. The additional control assigned to the AWP allows us to reduce the number of decision variables by removing the two height variables from the previous version and focusing exclusively on variables in the lateral directions. We investigated the response surface from the AWP compared to the common linear well description when the toe of a well was perturbed over a fault. In this test the response surface when using AWP was smoother than using the linear description. In general, a reduction in the number of variables will reduce the dimension of the search space, reducing the space to be explored by the optimization algorithms. To investigate the effect of reducing the dimension of the search problem through AWP, two different well placement scenarios were developed, optimizing the placement of one and three producers, respectively. In both scenarios, increased objective function values were observed using AWP. In addition, our methodology showed a generally higher convergence rate compared to the linear well description.
In the last part of this thesis, we present a preliminary study of the potential for using different information retrieval geometries and decision makers. Here we show that reinforcement learning can learn from information perpendicular to the current drilling direction of the virtual drilling procedure, without the need for the information that is in front of the current position. In this section, we also demonstrate that the problem can be formulated and solved as a optimization problem. By using the solutions from the individual optimizations, it is possible to use supervised learning to train an artificial neural network that can perform the same task with significantly lower computational cost. | en_US |
