Vous êtes sur la page 1sur 10
SPE 38441 Society of Petrcloum Engineers Reservoir Simulation: Past, Present, and Future 4J.W. Watts, SPE, Exxon Production Research Company Cenyiht197 Soca Palo Exner. ‘sper ead ttn 67S emt Sten Son ‘paper was sled he Pry Canin be 1997 SPE Reser Seusston ‘ovo hove tt been ovewed oye Socay @ Porcaum Egret dae cj o “Srechn ty nw suber ita mara epee, oe ro neat) rec ay Ste Socal Pain Ergrens oer. rember Papers pevwrine outro saat putesten oven ty Este Caren Ergonrs acre reson eaeon, oops of ay bar ot the pape ‘ners purposes wi the win ser the Sealy Patou Engrs [Senuted Bemananornrnce'n pet m terse e an tara far mee han 30 Soon autaione may tt oe, coped. "the ‘cas mat conan conpoicet Soren Renecron Tx same USA taotareacnSaae Abstract Reservoir simulation is a mature technology, and nearly all ‘major reservoir development decisions are based in some way on simulation results. Despite this maturity, the technology is changing rapidly. Itis important for both providers and users of reservoir simulation software to understand where this change is leading. This paper takes a long-term view of reservoir simulation, describing where it has been and where it is now. It closes with a prediction of what the reservoir simulation state of the art will be in 2007 and speculation ‘Fegarding certain aspects of simulation in 2017, Introduction Today, input from reservoir simulation is used in nearly all major reservoir development decisions. This has come about in part through technology improvements that make it easier to simulate reservoirs on one hand and possible to simulate ‘them more realistically on the other; however, although reservoir simulation has come a long way from its beginnings in the 1950's, substantial further improvement is needed, and this is stimulating continual change in how simulation is performed. Given that this change is occurring, both developers. and. users of simulation have an interest in understanding where it is leading. Obviously, developers of new simulation capabilities need this understanding in order to keep their products relevant and competitive. However, people that use simulation also need this understanding: how else can they be ‘confident that the organizations that provide their simulators are keeping up with advancing technology and moving in the right direction? 333 In order to understand where we arc going, it is helpful to know where we have been. Thus, this paper begins with a discussion of historical developments in reservoir simulation ‘Then it briefly describes the current state of the art in terms of how simulation is performed today. Finally, it closes with some general predictions. The Past This paper views the past as a progression, a series of advances. Sometimes this takes the form of a graph, such as CPU speed versus time. In other cases, i is a listing of events grouped by decade. In either case, the intent is to convey an impression of how rapidly development has occurred, with the hhope that that will assist in estimating the future rate of change. Computing. The earliest computers were little more than adding machines by today's standards. Even the fastest computers available in the 1970's and early 1980's were slower and had less memory than today's PC's. Without substantial progress in computing, progress in reservoir simulation would have been mesningless. Figure 1 gives the computing speed in millions of floating point operations per second for the three fastest CPU's of thei times: the Control Data 6600 in 1970, the Cray 1S in 1982, ‘and a single processor on a Cray T94 in 1996, The performance figures are taken from Dongarra's compilation of LINPACK results! and should be reasonably representative of reservoir simulation computations. Figure I shows single processor performance. Today, high-performance computing is achieved by using multiple processors in parallel. The resulting performance varies widely depending on problem size and the number of processors used. Use of, for example, 32 processors should lead to speedup factors of 15-25 in large reservoir models. As. a result, although Figure 1 shows rate of performance improvement to be slowing, if parallelization is factored in, it may actually have accelerated somewhat. No attempt is made to incorporate parallel performance into Figure 1 because of its wide variability. Consider Table 1, which comperes the Cray 1S to today's Intel Pentium Pro. The 1S was the second version of Cray's, 2 JW WATTS: SPE 32441 first supercomputer. Not ony was i the state of the art of its, time, but it represented a tremendous advance over its contemporaries. As such, it had the standard supercomputer price, a little under $20 million. It comprised large, heavy pieces of equipmem whose installation required extensive building and electrical modifications. The Pro, on the other hand, costs a few thousand dollars, can be purchased by mail order or from any computer store, and can be plugged into a simple wall outlet. The 200 MHz Pro is over twice as fast as the Cray (according to Dongarra’ numbers) and is commonly available with upto four times as much memory. Model Size, Over time, model size has grown with computing speed. Consider the maximum practical model size to be the largest (in terms of number of gridblocks) that could be used in routine simulation work. In 1960, given the computers available at the time, this maximum model size was probably about 200 gridblocks. By 1970, it had grown to about 2000 sridblocks. In 1983, it was 33,000; Exxon’ first application of the MARS program was on a model of this size running on the Cray 1S*, In 1997, it is roughly 500,000 gridblocks for simulation on a single processor. Figure 2 plots these values. This semi-log plot is a nearly straight line, indicating a fairly constant rate of growth in model size. The growth in model size is roughly consistent with the growth in computing speed shown in Figure 1 Technical Advances, The first "eservoir simulator" modeled single-phase flow in one dimension. Today, the norm is three- phase flow in three dimensions with many gridblocks and complex uid representation. Many advances in computational methods made this transition possible. Table 2 lists some of these by decade. In general, the advances in Table 2 are chosen because they are still in use today or they paved the way for methods used today. Also, they are methods that are used in many types of simulation, as opposed to techniques for modeling specific phenomena such as relative permeability hysteresis or flow in fractured reservoirs, The high points in the tble are discussed below. Reservoir simulation began in 1954 with the radial gas flow computations of Aronofsky and Jenkins’. The first work to receive notice outside the field of reservoir engineering was Peaceman and Rachford’s development of the alterating- direction implicit (ADI) procedure’. More than 40 years later, [ADI is still being used, though seldom in reservoir simulation Today, itis hard to appreciate how little was known atthe beginning of the 1960s. Even concepts that today seem obvious, such as upstream weighting, were topics of debate Yet, by the end of the decade, the first true general-purpose simulators had come into being, One of the difficulties in the 1960's was solving the matrix equations, Today, a solver must be fast and casy 10 use. Then, there were problems that could not be solved at all. The first effective solver was SIP‘. Though it was sometimes 334 troublesome to use, it nearly always could be made to work. ‘Another mathematical breakthrough of the time was development of implicit-in-time methods, which made it practical to solve high flow velocity problems such as well coning The 1970's saw publication of Stone's three-phase relative permeability models'!. These continue to be widely used. ‘Another innovation that has stood the test of time was the two- point upstream method". Despite widespread efforts since then, only incremental improvements to it have been found. Also having tremendous lasting impact was the development of solvers that used approximate factorizations accelerated by corthogonalization and minimization”. These made possible methods that were largely parameter-free. Finally, Peaceman’s well correction for determining bortor-hole pressure from gridblock pressure and well rate™ is almost universally used today In the early 1980's, a significant advance occurred with the development of nested factorization”. Both fast and very robust, nested factorization may be today’s most widely used matrix solver method. Another major step occurred in compositional simulation. Although the first compositional simulators were developed in the late 1960's, their formulations included an inherent inconsistency that hurt their performance”. The volume balance and Young- Stephenson” formulations solved this problem and, at the same time, made it practical to write a simulator that ean efficiently solve both black-oil and compositional problems. Development of cornerpoint geometry made it possible to use non-rectangular gridblocks”, providing a capability that was useful in a variety of applications. In the late 1980's, efforts shifted to issues related 10 geologic modeling, geostatistics, upscaling, flexible grids, and parallelization, and this emphasis has continued in the 1990's, The list of accomplishments for the 1990's is much shorter than those for the preceding three decades. This is, of course, partly because the decade is not finished yet, but it may also be partly because not enough time has elapsed to make lasting contributions recognizable. Perhaps it also stems in part from the diversion of effort from general computational work into the nuts and bolts of interactive software and parallelization Finally, it also may be, unfortunately, a result of the reductions in research effort that began in the mid-1980"s Simulator Capabilities. The preceding section discusses technical advances. They would not be of interest had they not led to improvements in capabilities from the user's standpoint, Table 3 lists, again by decade, the state of the art capabilities that were available in simulators of the time. No attempt is made to cite the literature in this discussion. These capabilities tended to become available in several simulators at about the same time, and offen there was no external publication describing them. ‘The computers of the 1950's permitted use of only the SPE 35441 RESERVOIR SIMULATION: PAST, PRESENT, AND FUTURE 3 ‘rudest models. Three-dimensional simulation was out of the question, and only small two-dimensional models were possible. Everything had to be kept simple, so single-phase flow or incompressible two-phase flow and very simple ‘geometry were used. ‘The more powerful computers of the 1960's enabled more realistic description of the reservoir and its contents. Three phase, black-oil fluid treatment became the norm. It became possible {0 run small three-dimensional models. Multiple ‘wells were allowed for, they could be located where desired, and their rates could be varied with time, Gridblock sizes could vary, and gridblocks could be "keyed out," or climinsted from the system. By the end of the decade, implicit computational methods were available, permitting practical well coning modeling. The 1970s seems to have been the enhanced oil recovery (GOR) decade. The first compositional simulators. were developed. Computing limitations forced these to use large ‘gridblocks, so numerical dispersion was a problem. Also, there were weaknesses in the initial formulations used. Nonetheless, they made it possible to begin to model phenomena that had up to then been ignored. Because of heavy interest in EOR, much effort went into modeling, miscible and chemical recovery. Finally, advances in implicit, computational methods provided the solution stability required to model thermal processes. In the 1980's, it became no longer adequate for the user to tell the simulator where to put the wells and how much to produce from them. In prediction runs, the simulator became responsible for making such decisions on its own. This led to development of complex well management software that, ‘among other things, determined when to work over existing wells or drill new ones, adjusted rates so as to adhere to constraints imposed by separator capacities, maintained ‘computed reservoir pressures at desired values, and allocated produced gas to injection and gas lift. Other developments led to approaches for modeling fractured reservoirs. The normal restriction of topologically rectangular grid connectivity was lifted to allow taking into account shifting of layers across faults. Finally, work began on interactive data preperation and display and on graphical user interfaces in general ‘The dominant efforts of the 1990's have been on various ‘ways to make simulators easier fo use. These have included continuation of the work on graphical user interfaces, ‘widespread attempts at data integration, and development of ‘automatic gridding packages. The use of numerical geologic models, often depicting fine-scale property variation, generated statistically, has become widespread. There has ‘been considerable work on methods for *upscaling" reservoir properties from these models’ cells to the much larger sridblocks that simulators can use. Gridding flexibility has increased through use of local grid refinement and more complex grid geometries. A current thrust is integration of simulation with non-reservoir computational tools such as facilities models and economics packages. Evolution of Software and Support. As simulators. became more powerful and flexible, their user communities changed. As this happened, developers and supporters had to change the way they did their jobs. As a result, software development and support practices progressed through several stages. The progression is still under way, with one stage left to be accomplished. 1. Developer use. Initially, a very small team of developers devised computational methods, implemented them in simulator, and applied this simuletor themselves. The first applications were intended to test the simulator and its concepts and the next ones to demonstrate its usefulness on real problems. After these tests, the developers made production runs intended to benefit the corporation. Portability of this simulator was not an issue, because there was only one computer for it to run 2, Team use. By the next stage, the team had grown and developed internal specialization, Within it, one group developed the simulator and another applied it.The simulator required customization foreach new application. Despite the group's specialization, developers were still frequently involved in applications because of the simulators need for continual support ‘The simulator frequently failed, and it was essentially undocumented. The simulator ran on a single computer, and portability was still not an issue. 3. Local use. In this stage, the simulator was used by people ‘who were located near to its developers but worked in other parts of the organization. The simulator still required customization for most new studies. Other support was requited frequently but not continually. Failures still occurred, but less frequently than before Documentation was adequate to permit use of the simulator with some assistance from developers. The simulator ran on several computers, but they were all of the same type 4, Widespread use. In this stage, the simulator first began to receive use by people at remote locations. It seldom required customization, but it still needed occasional support. It rarely failed. Documentation was thorough, but training was required for effective use of the simulator. Most applications were performed by specialists in the use of the simulator. The simulator ran ‘on a small variety of computers, 5. General use By this stage, the simulator will have become widely used by people with varying expertise. It rarely will need customization, will requite support only infrequently, and seldom will fail. Its documentation will be thorough and easily understood. Its user interfaces, will be intuitive and standardized. Little training will be 4 JW. WATTS: ‘SPE 30441 required to use the simulator. The user will need knowledge of reservoir engincering, but he will not need to be a simulation exper. Each transition to a new stage changes what is required of the simulator and those who support it, Each transition has been ‘more difficult than the ones that preceded it. A tra was surprisingly difficult was from local use to use, In the local use stage, the simulator was being used frequently and for the most part was functioning correctly. As a result, it seemed safe to send it to remote, perhaps overseas, locations. Doing so led to many more problems than expected. The remote computer differed slightly from the developer's, its operating system was at a different release level, and it was configured differently. The remote users tried the simulator once or twice, and it did not work. These users did not know the developers personally, and they had no really convenient way to communicate with thera. Typically the users abandoned attempts to use the simulator, and the developer was slow to lear of this failure. Interestingly, vendors were forced by their business to make the transition to widespread use before petroleum companies’ in-house developers had to. The vendors have been dealing with the related problems since the early 1970's; in-house developers began addressing them later, with varying degrees of success, This transition is made difficult by the high standards in usability, functionality, and robustness that, rust be met. Developing software and documentation of the quality needed is very time-consuming and requires different sets of skills than those traditional to reservoir simulator development ‘Vendor History and Role, Researchers working at major oil company laboratories developed the first reservoir simulators. It was not until the middle 1960's that vendors started to appear. Following is a brief history of certain of these vendors. Inclusion of a particular vendor in this discussion is nnot intended to imply endorsement, and exclusion is not intended to imply criticism. The firms discussed are those with which the author is familiar; for this reason North American-based firms are more likely to be included than those based elsewhere. The first to market a reservoir simulator was D. R. McCord and Associates in 1966, Shortly thereafter, Core Laboratories also had a reservoir simulator that they used in consulting work. ‘Tne year 1968 saw the founding of INTERCOMP and Scientific Software Corporation, two companies that dominated the reservoir simulation market in the 1970’s Despite their market success, a number of new companies were formed in the 1970°s and early 1980's, The first was INTERA, which was formed in 1973 by merging an INTERCOMP spinoff with the environmental firm ERA. INTERA initially focused on environmental work, but eventually got into reservoir simulation as discussed below. In 1977, the Computer Modelling Group was formed in Calgary with support from the province of Alberta. J. S. Nolen and Associates and Todd, Diettrich, and Chase were formed in 1979. In 1981, a reservoir simulation business was formed at the existing exploration-related firm Exploration Consultants Limited (ECL). Finally, SimTech was founded in 1982 and Reservoir Simulation Research Corporation in 1984 The founding of these firms was followed by a series of mergers and acquisitions. These began with in 1977 with the acquisition of INTERCOMP by Kaneb Services. In 1983, Kaneb Services sold INTERCOMP to Scientific Software Corporation, the two firms merging to form Scientific Software-Intercomp, oF SSI In the middle 1980°s, J. S. Nolen and Associates was acquired by what came to be Westem Arlas. In 1996, Landmark Graphics acquired Wester Atlas’ reservoir ulation business. Since then, Landmark was in tum acquired by and became a division of Halliburcon. In the middle 1980's, INTERA acquired ECL's reservoir simulation business. A few years later, INTERA split into two completely separate companies, one based in the United States and the other in Canada. The reservoir simulation business went with the Canadian INTERA. In 1995, Schlumberger’ GeoQuest subsidiary acquired INTERA's reservoir simulation business. Shortly later, the United States-based INTERA, which had become part of Duke Engineering and Services, reentered the reservoir simulation business by acquiring SimTech Finally, in 1996, the Norwegian firm Smedvig acquired Reservoir Simulation Research Corporation. AAs of the second half of the 1990's, vendor simulators are very widely used. As vendor products have improved, some petroleum companies have reduced or dropped altogether their in-house efforts. Those companies found that vendors could provide tools that were at least as,good as those that they could develop themselves, and that they could lower their development and support costs by using the vendors’ products. On the other hand, several large companies ‘continue to find it in their best interests to develop proprietary tools, perhaps incorporating certain vendor components into their systems. The Present The following discussion briefly reviews today's common practices in reservoir simulation. It is intended to depict what People actually do when performing state of the art simulation. It focuses on large, field-scale models. Geologic Modeling. The reservoir is defined using a numerical geologic model. Often the fine detail ofthis model is generated statistically. The mode! typically has on the order cof one to ten million cells. It is based on geologists’ beliefs about the reservoir’s depositional environments and data from a variety of sources such as seismic, well logs, and cores. At SPE 38461 RESERVOIR SIMULATION: PAST, PRESENT, AND FUTURE. 6 times, several randomized versions of the model, called “realizations,” are simulated in order to quantify uncertainty in the simulation results. Gridding. The user defines major reservoir units to which the simulation grid must conform, Within these units, the grid is, generated automatically with some guiding input from the user, The gridblocks are generally rectangular or nearly so. Local refinement of the grid in key parts of the model is used. Special computations account for flow across faults between eridblocks that are not opologically neighbors. Upscaling. The typical simulation gridblock consolidates several tens to several hundreds of geologic model cells. Effective permeabilities on the simulation grid are determined from the geologic model permeabilities using either an appropriate numerical average or flow-based upscaling ‘Sometimes finely-gridded segment models are run prior to field-scale simulation, with the results ofthese runs being used. to determine pseudo relative permeabilities, thus effectively accomplishing multiphase upscaling, Data Import, Well data exist in databases. Historical production and injection rates are exported from these databases in a form that can be read by the simulator or converted by utilities into simulator input. Most of this process is essentially automatic in the sense that it requires litle intervention by the user. On the other hand, the user ‘must work with historical pressure data and data describing the wells (such as completion intervals and tubing sizes) to get them into the form needed by the simulator. Fluid property data are also available in databases. Laboratory-measured results vary from sample to sample, however, and the user must manually average them in some way. Similar manual averaging is also needed for relative permeability and capillary pressure data. In addition, pseudofunctions are often used. Use of pseudofunctions is declining as finer grids become possible. Simulation, The following characterizes today’s reservoir simulation based on certain aspects of how it is performed Type. The large majority of simulations are of conventional recovery processes, and most of these simulations use black-oil fluid treatment with variable bubblepoint pressure. However, use of compositional fluid represeniation is growing and has become significant. When a compositional representation is used, fluid behavior is normally computed using an equation of state. Full compositional fluid representations are also used in simulating miscible recovery, particularly at laboratory scale. For fiekd-scale processes, the norm is to use simpler approaches, particularly in modeling carbon dioxide flooding. Simulation of steam injection processes is common, but simulation of in situ combustion is raze, partly because use of in situ combustion is rare and partly because simulating itis, difficult and expensive, Most simulation effort probably goes into modeling entire reservoirs or substantial portions of them, but in terms of number of models run there may be more simulations of single wells. This is being driven largely by the growth in use of horizontal wells. ‘Model Size. 100,000 gridblock field-scale black-oil models are common, and models several times larger are not ‘unheard of. Compositional models tend to be smaller by perhaps a factor of cree to five. Typical vertical-well coning models have about a thousand to several thousand gridblocks, while models of horizontal wells are a factor often larger. Most steam process models are ofa single well and have at most a few thousand gridblocks Computing Platform. Most reservoir simulation is performed on Unix workstations and servers. Simulation performed on other platforms is probably split fairly evenly between PC’s and supercomputers, Currently very little is, done in parallel, but that is changing rapidly. Simulation pre- and post-processing are done predominantly on Uni workstations, but use of high-end Windows-based PC's is, growing. History Matching. History matching is still essentially a manual process, assisted perhaps by software that organizes and displays results of history matches. Fully automatic history matching is still not feasible. Some optimization of small” sets of parameters, such as regional permeability multipliers, is performed automatically Predictive Well Management. During prediction, the simulator in effect operates the reservoir. It drills new wells, recompletes existing ones, restricts rates to satisfy facility constraints, switches wells to low pressure gathering systems, tc. It decides when to perform these actions based on rules and other information provided by the user. However, modeling of flow in the tubing and surface network is less sophisticated. Pressure drops are normally determined only in the wellbore, and these computations are typically made using tables. Most network computations are based entirely on material balance. In effect, physics is largely ignored after the fluid leaves the wellhead. This is changing, and some simulators can model these flows more rigorously. Use of Results. Reservoir simulation is most often used in making development decisions. These begin with initial evelopment and continue with decisions such as those relating to in-fill drilling and implementation of enhanced oil recovery processes. Simulation is not commonly used in day- to-day operations, but it is used at times in making major ‘operating decisions. 6 2.W. WATTS: ‘SPE 33441 ‘The Future Before attempting to predict the future, it is good to be aware of the business drivers, the most important of which are discussed below. These lead into current technical objectives, also discussed below. These objectives then form the basis of several predictions. Business Goals. In the view of petroleum company ‘management, research and development related to reserve simulation should address the four goals discussed below. ‘The order in which they are listed does not indicate priority, and which is most important varies from situation to situation. Compress calendar time. Time is money, and the more quickly a study can be performed, the more valuable it is Deadlines associated with possible property acquisitions are getting tighter, and competitive pressures are growing, In addition, a study that does not take long probably will not cost ‘much, Decrease cost. Petroleum companies continue to be under pressure to reduce costs. Despite reservoir simulation’s computing-intensive nature, its major cost comes from the engineering labor it requires. ‘Reduce expertise required. In a sense, this goal is closely related to the preceding one, since maintaining expertise is expensive. However, it also relates to the widespread push to decentralize and flatten organizations. As this occurs, more simulation is being done locally by generalist reservoir engineers, rather than by centralized groups of specialists ‘These reservoir engineers cannot be expected to be simulation experts Improve accuracy and realism. These are the traditional goals of simulation research and development. New challenges have been created by the grids coming out of geologic models and by more complex well geometries. ‘Technical Objectives and Efforts to Address Them. A good technical objective must meet two criteria: it must address one ‘or more of the business goals, and it must be achievable with, reasonable effort in the intended time frame. Substantial progress on the following objectives should be possible within 10 years, and achievement of them should be possible within 20 years. The following states each objective, expands upon it very briefly, and describes current work addressing it. Require’ the user to provide only data describing the physical system, its lids, and its history. Do not require him to specify computational data such as solver selection, iteration parameters, and timestep controls. Create the simulation grid for him with no intervention on his par. Several organizations are working on gridding, and some of their work relates to automating the gridding process. ‘Current linear equation solver work relates primarily to parallelization, but within this work there is an ongoing fitempt to improve robusiness and reduce the effort required by the user. 338 Automate history matching. Determine the best possible fit to existing historical data, consistent with the presumed depositional environment and other characteristics of the geologic model Recent work has led to an economical way to compute derivatives for use in gradient-based optimization ‘methods. These will become more commonly applied, but ‘much more is needed for truly automatic history ‘matching. Minimize the time and effort required to access information required to perform the study and to generate the results that are the study's objective. Integrate data when doing so is practical and provide efficient import capa when it is not. Likewise, integrate downstream computations ‘when practical and provide efficient export capabilities when rot. The Petroleum Open Systems Corporation (POSC) and its members are addressing the data integration issue, Recent acquisitions of reservoir simulation vendors by larger service organizations is leading to integration with computing tools both upstream and downstream of simulation. Predictions, Following are predictions regarding the year 2007 state of the art of reservoir simulation and related technologies. 1. The dominant high-end computing platform for simulation calculations will be a Unix server comprising multiple nodes, with each node having @ small number of processors. Where high performance is not needed, the dominant platform will be a ‘multiprocessor PC running NT. 2. The dominant pre- and post-processing and visualization platform will be the top of the line version cof whatever the PC has become, 3. Integration of reservoir simulators with the software and databases that provide their input data will be much better than itis today. 4. Reservoir simulation will be essentially automatic, given the geologic model, fluid data, and rock (ie, relative permeability and capillary pressure) data 5. The largest black-oil simulations witl use at least 10 million gridblocks. 6. Most simulators will be based on unstructured grids. 7. Integrated reservoir-surface network calculations will bbe common, 8. Use of history maiching tools will be widespread, but the history matching process will not be automatic ‘The above predictions are made with some confidence. It seems reasonable to expect most of them to be generally correct, with perhaps one or two tuming out wrong Attempting to predict further into the future is more problematic. Nonetheless, it may be instructive to consider

Vous aimerez peut-être aussi