Uncertainty Analysis of Otway CO2 Sequestration Project

A viable means of CO2 reduction in the atmosphere is to capture and concentrate CO2 from large point sources such as power plants and petroleum refineries and store it by underground injection. The main concern for commercial scale CO2 sequestration in geologic formations is the uncertainties associated with this process. The risks involved in different stages of a CO2 sequestration project are related to geological and operational uncertainties. This paper presents the application of a grid-based Surrogate Reservoir Model (SRM) to a real case CO2 sequestration project in which CO2 were injected into a depleted gas reservoir. This project is a part of the National Risk Assessment Partnership (NRAP)’s research on CO2 sequestration process. An SRM is a customized model that accurately mimics reservoir simulation behavior by using Artificial Intelligence & Data Mining techniques. Initial steps for developing the SRM included constructing a reservoir simulation model with a commercial software, history matching the model with available field data and then running the model under different operational scenarios or/and different geological realizations. The process was followed by extracting some static and dynamic data from a handful of simulation runs to construct a spatio-temporal database that is representative of the process being modeled. Finally, the SRM was trained, calibrated, and validated. The most widely used Quantitative Risk Analysis (QRA) techniques, such as Monte Carlo simulation, require thousands of simulation runs to effectively perform the uncertainty analysis and subsequently risk assessment of a project. Performing a comprehensive risk analysis that requires several thousands of simulation runs becomes impractical when the time required for a single simulation run (especially in a geologically complex reservoir) exceeds only a few minutes. Making use of surrogate reservoir models (SRMs) can make this process practical since thousands of SRM runs can be performed in minutes. Using this Surrogate Reservoir Model enables us to predict the pressure, phase saturation, and CO2 distribution throughout the reservoir with a reasonable accuracy in seconds. Consequently, uncertainties associated with reservoir characteristics and operational constraints are analyzed with the SRM in a considerably shorter amount of time compared to the conventional uncertainty analysis techniques. Surrogate reservoir modeling opens new doors in reservoir modeling by providing the means for extended study of reservoir behavior with very low computational cost. Introduction Despite all the efforts in shifting the energy sources to the renewable and atmosphere friendly source of energy, fossil fuels are still the most essential source of energy for industries and transportation. Considering the demand growth it is believed that fossil fuel consumption will continue to increase through the next century. As a result, concerns about the greenhouse gas emission and its impact on global warming and climate change are increasing. This has encouraged focus on two different approaches of reducing CO2 in the atmosphere. The first one is the preventive methods which aim at 2 minimizing CO2 emission in to the atmosphere through improved efficiency, renewable energy supplies, carbon-free fuel consumption and nuclear fission, and the second approach is to apply the remedial methods through which the CO2 concentration in the atmosphere is reduced [1,2]. A viable means of CO2 reduction in the atmosphere is to capture and concentrate CO2 from large point sources such as power plants and petroleum refineries and store it by underground injection. The process is called Geological Carbon Sequestration (GCS). Three types of target reservoirs are capable of sequestering large volumes of CO2 which are depleted oil and gas reservoirs, saline aquifers and coal beds. In these reservoirs, high pressure and temperature make CO2 a supercritical state in which CO2 acts as a liquid. Therefore, it would be possible to inject a large volume of CO2 into a limited pore volume of reservoir [3]. The main concern of commercial scale CO2 sequestration in geologic formations is the uncertainties associated with this process which is directly related to the geological and operational uncertainties of the process. Since the CO2 is injected underground there is no direct method to determine the CO2 flow in the porous media. The reservoir simulation models are the only tools through which the CO2 fate can be studied. These models are constructed based on the geological studies and interpretations, field observation and measurements and therefore are essentially uncertain. In each specific sequestration project, different operational practices will have different sequestration outcomes. Consequently, any practical uncertainty analysis and risk assessment technique should address both geological and operational uncertainties. The most widely used uncertainty analysis techniques, such as Monte Carlo simulation, require thousands of simulation runs for implementing a comprehensive uncertainty analysis for a specific CO2 sequestration project. In creating the reservoir simulation models on one hand adding complexity to the reservoir simulation model is inevitable since integrating all the observations and measurements is the sensible way to reduce the uncertainty and on the other hand the more complex the simulation model, the higher the run time. Therefore, using the conventional geo-statistical approach for uncertainty analysis of a fairly complex reservoir needs thousands of simulation runs makes it impractical. Other conventional method that is currently used for uncertainty analysis is generating a Response Surface for the problem. Response Surface Methodology (RSM) is a collection of statistical and mathematical techniques useful for developing, improving, and optimizing processes [4]. The field of response surface methodology consists of the experimental strategy for exploring the space of the process or independent variables, empirical statistical modeling. It develops an appropriate approximating relationship between the output and the process inputs [5]. In order to develop a Response Surface for uncertainty analysis, several combinations of the input parameters (chosen from a reasonable range) are generated and hundreds of simulation runs outputs are provided through making different realizations in a reservoir simulator. Using the outputs from simulation runs and based on some mathematical methods, a surface is generated to represent the possible responses which is resulted from the predetermined realizations. Different realizations are selected so that maximum coverage is obtained with minimum simulation runs. Some other techniques are used such as Latin Hyper Cube and Design of Experiments in order to optimize this process [6, 7]. Generally, the geo-statistical approaches such as those explained above suffer from major shortcomings such as: 1. They still need hundreds of simulation runs to be capable of effectively perform a comprehensive uncertainty analysis. 2. Once the simulation runs for all realizations were completed, the response surface is generated 3 using the outputs and thus the input parameters no longer play any role. In other words, it would not be possible to evaluate the output if any of the inputs are to be changed to the value that does not exist in the developed realization. In this study, we are making use of the Surrogate Reservoir Modeling (SRM) technique [8] for uncertainty analysis of a real CO2 sequestration project. This technique is not based on statistics and therefore just needs a handful of reservoir simulation runs in order to perform the required analysis. SRM is a customized model that mimics reservoir simulation results by using Artificial Intelligent & Data Mining techniques with a very low computational cost. It consists of one or several neural networks which are trained, calibrated and verified using very small portion of data. SRM can be constructed as well-based or grid-based. Well-based SRM is able to make predictions for the well parameters such as rate of oil, gas and water production [9, 10] and the grid-based SRM will return the grid level parameters such as pressure and liquid/gas phase saturation. One of the other advantages of SRM is that once an SRM is created for a specific problem, the input parameters can be changed (within a range) at any time and the output response to this change can be observed and evaluated. Moreover, the impact of any input alteration on the output is obtained within a few minutes regardless of the complexity of the reservoir model [9].