Optimization problems, such as optimal well-spacing or completion design, can be resolved rapidly via surrogate proxy models, and these models can be built using either data-based or physics-based methods. Each approach has its strengths and weaknesses with respect to management of uncertainty, data quality or validation. This paper explores how data- and physics-based proxy models can be used together to create a workflow that combines the strengths of each approach and delivers an improved representation of the overall system. This paper presents use cases that display reduced simulation computational costs and/or reduced uncertainty in the outcomes of the models. A Bayesian calibration technique is used to improve predictability by combining numerical simulations with data regressions. Discrepancies between observations and surrogate outcomes are then observed to calibrate the model and improve the prediction quality and further reduce uncertainty. Furthermore, Gaussian Process Regression is used to locate global minima/maxima, with a minimal number of samples. To demonstrate the methodology, a reservoir model involving two wells in a drill space unit (DSU) in the Bakken Formation was constructed using publicly available data. This reservoir model was tuned by history matching the production data for the two wells. A data-based regression model was constructed based on machine learning technologies using the same dataset. Both models were coupled in a system to build a hybrid model to test the proposed process of data and physics coupling for completion optimization and uncertainty reduction. Subsequently, Gaussian Process Model was used to explore optimization scenarios outside of the data region of confidence and to exploit the hybrid model to further reduce uncertainty and prediction. Overall, both the computation time to identify optimal completion scenarios and uncertainty were reduced. This technique creates a robust framework to improve operational efficiency and drive completion optimization in an optimal timeframe. The hybrid modeling workflow has also been piloted in other applications such as completion design, well placement and optimization, parent-child well interference analysis, and well performance analysis.

Computer vision (CV) techniques were applied to X-ray computed tomographic (CT) images of reservoir cores to evaluate their potential for rapidly identifying fractures and other lithologic/geologic characteristics. The analyses utilized feature-labeled CT cross sectional images, themselves distributed submillimeter voxels of density- and atomic number-sensitive CT Numbers, as inputs to a Fully Convolutional Neural Network (FCN) for semantic segmentation of reservoir core features. In FCN, an image is interrogated using a series of sliding windows at various scales to create weighted filters to reduce error between classes in training images. These networks of filter layers were used to assign probabilities of classes, which were upscaled back to the original image dimensions resulting in probabilistic class assignments onto each pixel. FCN model accuracy, defined by its ability to replicate manually-assigned labels in the raw (unannotated) training image stack, was at least 80% and generally improved with the size of the training set. Once the labels were assigned, the underlying feature frequency, orientation, and size were measured in 3D volume reconstructions using algorithms modified from standard image analysis software. This method allowed users to endow a classification model with subject matter knowledge for further, autonomous label prediction. Thus, while initial image annotation was labor intensive, subsequent images were rapidly classified once the model was built. The classified labels were analyzed for abundance, orientation, and size of fractures were calculated to characterize spatial information of these features. FCN combined with fracture labeling improved knowledge capture and automation of fracture identification. Models trained by high quality 3D datasets can greatly reduce the time needed to describe subsequent core. The method demonstrated is not limited to fractures, other lithologic/geologic features could be trained using the same method, which may result in additional efficiencies.

Gas lift is one of the most widely used artificial lift methods, and the use of nodal analysis to generate the gas lift performance curve is well established. However, the optimal gas injection rate is often selected as the point with maximum liquid production, which neglects the cost of incremental injection gas volume. This paper investigates the determination of the optimal operational point using a multiobjective optimization technique by considering the trade-off between gas consumption and oil production. The indicator-based evolutionary algorithm transforms the multiobjective problem into a single objective one using the hypervolume metric computed in the objective space. For the gas lift problem, which is a bi-objective problem aimed at maximizing oil production while minimizing gas injection rate, the hypervolume metrics are identically equivalent to geometric hyperareas under the trade-off curve. The optimization is only applied to the monotonically increasing portion of the gas lift performance curve; thus, all trivial sub-optimal conditions are excluded. The optimal operational point of gas injection rate is determined by finding the maximum rectangular hyperarea under the performance curve. The proper determination of the optimal injection gas rate could not only improve the efficiency of the gas lift itself, but also reduce the burden on the maintenance of surface facilities. The method is also applied to the multi-well scenario where a novel multi-well gas lift performance curve is generated using multiobjective Genetic Algorithm, which could help determine the optimal gas allocation/distribution scenario. The described process is incorporated in an integrated workflow which further leads to fast delivery of analysis/results that enable production engineers to make smarter decisions faster in a repeatable way.

Due to complex characteristics of shale reservoirs, data-driven techniques offer fast and practical solutions in optimization and better management of shale assets. Developments in data-driven techniques enable robust analysis of not only the primary depletion mechanisms, but also the enhanced oil recovery in unconventionals such as natural gas injection. This study provides a comprehensive background on application of data-driven methods in the O&G industry, the process, methodology and learnings along with examples of data-driven analysis of natural gas injection in shale oil reservoirs through the use of publicly-available data. Data is obtained and organized. Patterns in production data are analyzed using data-driven methods to understand key parameters in the recovery process as well as the optimum operational strategies to improve recovery. The complete process is illustrated step-by-step for clarity and to serve as a practical guide for readers. This study also provides information on what other alternative physics-based evaluation methods will be able to offer in the current conditions of data availability and the understanding of physics of recovery in shale oil assets together with the comparison of outcomes of those methods with respect to the data-driven methods. Thereby, a thorough comparison of physics-based and data-driven methods, their advantages, drawbacks and challenges are provided. It has been observed that data organization and filtering take significant time before application of the actual data-driven method, yet data-driven methods serve as a practical solution in fields that are mature enough to bear data for analysis as long as the methodology is carefully applied. The advantages, challenges and associated risks of using data-driven methods are also included. The results of data-driven methods illustrate the advantages and disadvantages of the methods and a guideline for when to use what kind of strategy and evaluation in an asset. A comprehensive understanding of the interactions between key components of the formation and the way various elements of an EOR process impact these interactions, is of paramount importance. Among the few existing studies on the use of data-driven method for natural gas injection in shale oil, a comparative approach including the physics-based methods is included but they lack the interrelationship between physics-based and data-driven methods as a complementary and a competitor within the era of rise of unconventionals. This study closes the gap and serves as an up-to-date reference for industry professionals.

In response to the low hydrocarbon price environment, most operators have decreased the number of rigs actively drilling wells ( Cui 2016 , Chapman et al. 2016 ). However, the production decline in wells completed in a given timeframe has undergone far more change than the count of active rigs. In many cases, wells must be drilled to meet contractual lease obligations but are not being completed due to uncertain economic conditions. Type curves are valuable tools to predict future performance of such drilled-but-uncompleted (DUC) wells by analyzing results of existing wells exhibiting similar characteristics. It is reasonable to assume, while considering reservoir characterization and completion engineering, that existing wells can predict the performance of future wells. However, to generate a larger sample size with which to construct the type curves, data can be normalized by a particular variable. A normalized curve is then scaled by an adjustment associated with the new well to produce its expected production profile. This same process also allows type curves to be applied to other areas with different reservoir properties or completion parameters. Analysis focuses on applying normalized type curves to estimate DUC well performance and populate a ranking system to guide the completion planning process. Reservoir material balance and surface network simulation are also utilized to understand subsurface effects and identify possible facility constraints above ground. Additionally, there is a brief discussion regarding type well best practices to aid in the analysis.

Abstract Electric submersible pumps (ESPs) and progressive cavity pumps (PCPs) are used as lifting methods in many oil fields to improve oil production. Downhole gauges are now installed on most of the ESPs and PCPs, providing relevant pressure and temperature measurements to assess pump performance and potential failures. The volume of high-frequency data overwhelms the capacity of most direct assessment methods based on human observation. To address this challenge, a method using a combination of signal processing tools together with an expert system has been developed to deliver a predictive diagnostic in real time. The data acquired is fed into a signal processing module to analyze and highlight trends and patterns. This module, based on Bayesian theory, segments the signal into a series of homogenous linear models. The segmentation is then used to provide statistical information about downhole measurements. Signal processing outputs have been designed to be probabilistic to account for uncertainties. The probability that pump intake pressure and pump discharge pressure will converge toward the same value is one example of such output. The probability of a noisy signal is another illustration of a result from the processing module. This real-time process is used to analyze all the ESP or PCP measurements available on specific criteria. Those pieces of information allow the expert system to predict pump failures. A Bayesian network with its probabilistic reasoning capabilities has been chosen to reproduce the ESP or the PCP physical knowledge. This Bayesian network contains input nodes gathering artificial lift parameters features and output nodes inferring failure probabilities. During a test in real conditions, this concept showed its capability to predict and correctly diagnose the failure. In addition, this solution confirmed its robustness with accurate failure prediction under numerous operational conditions—even when downhole parameters were missing. This concept could be used to reduce reaction time to a failure, improve the monitoring capacity of each surveillance engineer, and improve the ESP life expectancy by early remediation of failure signs.

Abstract Differential pipe sticking has a major effect on drilling efficiency and well costs. This problem is affected by many parameters, such as drilling fluid properties and the characteristics of the mud cake that is formed while drilling. Comparatively little research has assessed the prediction of stuck pipe. Traditionally, stuck pipe problems are solved by using some standard methods and techniques after they occur, but the real key to savings and success is in the avoidance of the risks associated with the stuck pipe. If these risks are identified in advance, better solutions can be provided to reduce the associated costs. To account for all of the aspects of differential pipe sticking and anticipated nonlinear behavior of the variables involved, fuzzy logic and neural network modeling can be used as primary predictive tools. These methods are widely used in other industries and in the petroleum field, especially in reservoir and core analyses. This paper presents a study of the application of the concepts of fuzzy logic to the problem of differentially stuck pipe. These methods make it possible to estimate the risk of stuck pipe occurrence in the well planning procedure and during drilling in real time. The nonconvolutional analysis of the model is based on the constraints of different drilling variables. Discriminant function analysis and fuzzy logic were used to classify drilling and mud variables. Extensive simulations were performed. This paper presents a case study in which the fuzzy logic and neural network modelling are successfully used to estimate and predict pipe sticking. Introduction Stuck pipe mechanisms are grouped as either mechanical or differential. Differential pipe sticking occurs when a part of the drill string casing or logging tool becomes embedded in a mud solid filter cake and is held there by a significant amount of differential pressure, as shown in Fig. 1. Mechanical pipe sticking occurs while drilling and is identified when the bottomhole assembly (BHA) cannot be rotated or moved because of junk in the hole. The proportion of incidents classified in each category varies with the type of well and the geographical area. Stuck pipe incidents have been one of the major technical challenges of the drilling industry and events typically result in a significant amount of downtime and remedial costs. The recent increase in drilling activity, shortage of experienced personnel and equipment, and drilling in higher-risks areas have increased the risk of stuck pipe events in all drilling operations (Yarim et al. 2007). Fundamental Parameters Affecting Differential Sticking There are many parameters that affect differential sticking. In this study, the general properties of the drilling fluids and mud cake will be reviewed to provide an understanding of how they affect the potential of pipe sticking occurrence and the necessary force to pull the pipe free after it is stuck. Mud Type. A comparison of generic mud types has shown oil-based muds to have the lowest stickance values and gel-waterbased mud has the highest. Polymer-water-based muds fall between these two extremes. It was found that the sticking potential also varies greatly within a mud type, depending on the precise formulation tested (Reid, Meeten, and Way 2000). Lubricant. The addition of certain lubricants for water- and oil-based muds will reduce the effect of differential sticking. If sticking still occurs, then reduce the force needed to free the stuck pipe or tool.

Abstract Minimizing production of water from gas reservoirs is one of the main strategies for enhancing primary hydrocarbon production. Advances in intelligent well technology and simulation of reservoir-production system enable optimum inflow allocation of produced fluids by controlling perforations and valve settings. However, real-time optimum control of flow is still a challenge. To this end, we have developed an advanced method for optimization of production based on a feed-back optimum control concept. The method allows optimization of inflow-control-device operation in conjunction with strategic updating of reservoir-smart well model under uncertainty. In this work, we present a hybrid optimization method for improving optimum solutions. In this method, the response of production system to control variables is mathematically described by a high-order proxy model, which is developed using Response Surface Methodology (RSM) and Design of Experiments (DOE). The validity of this method was tested for the dynamic optimum control of gas and water coning observed in a physical two-dimensional layered bead-pack model with automatic inflow control valves. The valves were actuated by distributed flow or pressure sensors. Results showed that dynamic (time-dependent valve settings) and static (fixed valve setting) control exhibited advantages one over another, depending on operational conditions. It was also found that the heterogeneity of porous media strongly influenced the control effectiveness. Field implications are discussed. Introduction For water-drive gas reservoirs, a well is conventionally perforated at the very top of the interval and produced at a restricted rate. However, this perforation scheme limits gas production rate and thus lowers gas recovery. Moreover, premature water breakthrough into the wellbore results in an increase in water-to-gas ratio (WGR). A well producing at high WGR, eventually, must be shut in because of the un-economical operation. Intelligent well completions with downhole inflow-control valves (ICVs) and distributed pressure and temperature sensors enable the control of inflow profile across the entire perforated wellbore. An optimum inflow-control strategy is further needed to enhance the selectivity of produced fluids through modification to the fluid distribution in the reservoir, particularly near the wellbore region. Several authors have demonstrated the applications and benefits of smart wells equipped with ICVs. Yeten et al.1 applied a conjugate gradient optimization technique to control intelligent multilateral wells. Brower and Jansen2 developed an algorithm to optimize ICVs setting for water flooding using optimal control theory. Gai3 applied modified inflow performance relationships to optimize valve settings for multi-zone completions. Ajayi and Konopczynski4 used a gradient-based optimization to optimized commingled production from a multi-zone intelligent completion. Naus5 employed sequential linear programming to optimize the operation of infinitely variable ICVs during commingled production. Although several studies have been done to optimize the ICVs, most of these optimization methods are gradient-based which require derivative information of computationally expensive reservoir process models. In addition, there is no guarantee that the solution based on gradient based schemes would be globally optimal. In this regard, we developed a hybrid optimization approach that combines both Genetic Algorithms (GA) (non gradient-based method) and steepest descent (gradient-based) method. This approach is used in conjunction with a proxy model to obtain optimum control of unwanted fluid production that can be proactively implemented. The proxy model is obtained by using experimental design and response surface method. Genetic algorithms have found several applications in the petroleum industry owning to their ability to handle highly nonlinear processes and obtain globally optimum solution (e.g. the cyclic steam scheduling problem6, well placement7, the design of gas production system8 and gas lift production9). However, application of a hybrid approach for dynamic optimum control of water coning in both laboratory-scale and field-scale implementations for gas reservoirs is not found in the literature.

Abstract Accurate information on phase behavior and properties of fluids is an essential element in proper management of petroleum reservoirs. These fluid properties which are usually determined by laboratory experiments performed on samples of actual reservoir fluid or using empirically derived correlations provide the information required to properly understand the phase behavior, evaluate various production scenarios, optimize reservoir production and IOR schemes, and to maximize ultimate recovery and optimize production economics. One of these properties is the petroleum reservoir fluid viscosity. Crude oil viscosity is an important physical property that controls and influences the flow of oil through porous media and pipes. This paper introduces a new application of fuzzy logic and neural networks in petroleum engineering. Artificial intelligence techniques such as neural networks, fuzzy logic and genetic algorithms for data analysis and interpretation are an increasingly powerful and reliable tool for making breakthroughs in the science and engineering and it is becoming clear that our industry has realized the immense potential offered by intelligent systems. The introduced model in this paper can predict the reservoir fluid viscosity data with neural networks and fuzzy logic approach. We can use these techniques in order to recognize the pattern between the given data sets where this pattern may not be understood clearly or no precise mathematical relationship exists Prediction of the proposed model has been tested against the measured reservoir fluid viscosity data. Results indicate that the proposed prediction model with recognizing the possible patterns between input and output variables can successfully predict and model reservoir fluid viscosity. Introduction Modern reservoir engineering practices require accurate information on thermodynamic and transport fluid properties together with reservoir rock properties to perform material balance calculations. These calculations lead to the determination (estimation) of the initial hydrocarbons in place, the future reservoir performance, optimal exploration and production schemes, and the ultimate hydrocarbon recovery. Crude oil viscosity is an important physical property that controls and influences the flow of oil through porous media and pipes. The viscosity, in general, is defined as the internal resistance of the fluid to flow. The oil viscosity is a strong function of the temperature, pressure, oil gravity, gas gravity and gas solubility. Whenever possible, oil viscosity should be determined by laboratory measurements at reservoir temperature and pressure. The viscosity is usually reported in standard PVT analyses. If such laboratory data are not available, engineers may refer to published correlations, which usually vary in complexity and accuracy depending upon the available data on the crude oil. The viscosity of crude oils is a critical property in predicting oil recovery. Viscosity reduction and thermal expansion are the key properties to increase productivity of heavy oils. Reservoir simulators are routinely used to predict and optimize oil recovery from oil fields. These simulators require as input properties of the reservoir fluids as a function of pressure, temperature and composition. The accuracy of the fluid properties can decisively affect the results of the simulation. Among the required fluid properties are phase densities, phase viscosities, formation volume factors and dissolved gas-oil ratios. The physicochemical properties of the reservoir fluids are a function of the fluids' composition. These compositions can be determined by experimental analysis such as, true boiling point essays and gas chromatography. In many practical cases no compositional information is present. A practical method to predict reservoir fluids' viscosities should be able to calculate viscosity of compositional and black oils. Numerous viscosity-correlation methods have been proposed. None, however, has been used as a standard method in the oil industry. Since the crude oil composition is complex and often undefined, many viscosity estimation methods are geographically dependent. Most correlation methods can be categorized either a black oil or as compositional.

Abstract Neural networks are a well developed computational intelligence technology with many possible applications in the oil and gas industry. For this study a neural network was developed to predict production potential for a single formation, prior to drilling, over a 16,000 square mile area of SE New Mexico. The process involved gathering data for use as potential inputs, collecting production data at known wells, selecting optimal inputs, developing and testing various network architectures, making predictions, analyzing and applying the results. Variables considered as potential inputs included the thickness of the primary source rock (Woodford shale), total organic carbon (TOC), production index (PI), paleo thickness, curvature of paleo structure, and permeability. Two variables were derived; the product of Woodford thickness, TOC and PI, termed generative potential, and the log of generative potential. Variables were interpolated using kriging over the study area. The neural network was trained to identify the production at a set of wells with attempted production from the formation. Production was measured in barrels of oil equivalent per month (BOEPM) averaged over the first year 12 months of production. The data included producing wells and dry holes. Potential input variables were tested against the output to determine the best variables for training. First, a fuzzy ranking algorithm selected Woodford thickness and TOC. A secondary correlation study also recommended Woodford thickness and TOC, and added permeability and generative potential. A series of neural networks were developed to test different architectures. The architectures were compared based on the r2 correlation while considering the possibility of overtraining based on the size of the input variable set. Random subsets of data were excluded for blind testing. A 4–6-1 architecture with an r2 correlation of 0.84 was selected. The neural network was re-trained multiple times achieving consistent correlations in each test. Once trained, the network was used to predict production over the entire region. This predicted production was further refined by excluding production at locations where the Woodford shale was not present. Results were evaluated by inspecting a map of predicted production and performing statistical testing, including a correlation of predicted and actual production, which produced a correlation coefficient of 0.79. The results were then used by the Devonian FEE Tool, an expert system designed to reduce exploration risk. Introduction An expert system that uses fuzzy reasoning, the Fuzzy Expert Exploration Tool, (FEE Tool) has been developed for the Devonian Carbonates of SE New Mexico by the Reservoir Evaluation and Advanced Computational Techniques (REACT) group at the Petroleum Recovery Research Center (PRRC) of New Mexico Tech. This tool was modeled on a previous successful expert system developed for the Lower Brushy Canyon formation1. The FEE Tool is built around a knowledge base of rules generated through interviews with experts having knowledge of the target formation. Each rule requires an input, either provided by the user or contained in a project database. As with the Lower Brushy Canyon expert system, a desired input value for the Devonian FEE Tool is predicted production. To obtain values of predicted production to use as inputs to the knowledge base rules, neural networks were used. The neural networks provided values of predicted production for each gridpoint of the 16,000 square mile region (one grid square equals 160 acres, with a total of 64,347 points for the region). To generate relevant values of predicted production for the region, data were collected for use as potential inputs and outputs for training a neural network. Potential inputs included geophysical data, structure data, source rock data and log data. Output data came from production reports for both producing and non-producing wells. For the purposes of this work, production was measured in barrels of oil equivalents or BOE, setting 6 mcf of gas equal to 1 BOE.

Abstract This paper is related to neural network processing of near well bore geological logging data. The paper has been prepared to address the increased use and interest in artificial intelligence technologies in the interpretation of open-hole and/or cased-hole formation evaluation data. A typical neural network model is developed from a training well or wells where measurements are available. The model is tested on a validation well or wells before it is applied to a specific field /reservoir. Quite often models are developed with limited data: for example single well data should be retrained with multi well data to improve its scope and applicability. Unfortunately, some efforts to upgrade can result in choosing to use sub-optimal training algorithms because of the need to deal with an extensively increased data size. The consequence is a reduction in predictive accuracy. To improve accuracy, it is wise to have ways remove faulty, redundant and insignificant data, detect inconsistent data, have the ability to "add", i.e., duplicate samples in key target zones. This paper discusses how cluster analysis can be integrated with graphic visualization and fuzzy decision making to support training-data selection in field model development based on multiple well data. A case study using cased hole logging data from a pulsed neutron tool demonstrates the method and its application in predicting the response of open hole triple-combo logs (neutron porosity, formation density and deep resistivity). Introduction Different predictive tools have been used in processing geological logging data for many years. A field data based predictive model usually takes selected measurements of specific logging tools as inputs and produces predicted outputs using either a deterministic function or an empirical function generated from a training process. As a typical predictive framework, an artificial neural network (ANN) 1 receives special interest and demonstrates increased use in petrophysical applications. To build an ANN model, data are selected from well logs, trained with optimization algorithms, and tested in different wells for validation. In the course of this process, data selection not only produces the greatest impact on the scope and applicability of the model, but also affects its accuracy and generalization performance. This statement is especially true if a single model for the field/reservoir is desired, and the data for all training wells and testing wells need to be normalized to a "field histogram" 2,3,4 . Since the uncertainty induced by different environmental factors and/or systematic errors may somehow corrupt the field data integration and pre-processing, special attention and treatment should be given to training-data selection. The training-data selection is more heuristic than systematic in most neural network applications. One of the common heuristic approaches is to use a predetermined data percentage to randomly select the training, validation and testing data sets, which may cause the training results to be sensitive to the specific data splitting if only single well data is available. For multiple-well training-data selection, it is quite often the case to define a resampling strategy to remove a certain amount of data in each individual well, and make the combined data set fall within a specific size limit. This procedure will allow the use of some powerful, but memory-constrained training algorithms (Levenberg-Marquardt-based algorithms 5 , for example). Otherwise, some sub-optimal training algorithms (gradient-descent-based algorithms) must be used with sacrificed training accuracy. However, as discussed above, decision-making is difficult in determining the resampling strategy without a deep understanding of the nature of the multiple well data. Evenly scattered interval sampling (systematic sampling with respect to depth) with reduced density may remove some redundant data, but may also remove some useful information at the same time such as thin bed data.

This paper was prepared for presentation at the 1999 SPE Mid-Continent Operations Symposium held in Oklahoma City, Oklahoma, 28-31 March 1999.

This paper was prepared for presentation at the 1999 SPE Mid-Continent Operations Symposium held in Oklahoma City, Oklahoma, 28-31 March 1999.

Abstract Recent years have seen a growing trend towards "right sizing" in the oil business. This has lead to improved efficiency and greater technical challenges. One of the off spins of this process is the current trend to limit up front investments in processing capacity, by designing it to the minimum production stream that will yield the maximum profit. This paper addresses these challenges and describes a process of combining 3D simulation results and Linear Programming (LP) to optimise asset value under numerous production and drilling constraints. The subsequent application maximises post tax NPV, and can be used for optimising drilling schedules, IOR projects, and process upgrades simultaneously. A newly developed well production rate formulation, including influence functions, allows for large time savings and a robust solution. This formulation also allows choking of wells without loss of reserves. This work differs from preceding studies by allowing significant reduction in data requirements, and incorporating an iterative solution between LP and reservoir simulation. This also opens the possibility to verify the accuracy of the optimal solution. The number of simulation runs required is reduced by a factor 40 compared to conventional LP formulations. The Ekofisk area in the North Sea is used as an example of the application. This area comprises several fields and platforms and is currently facing huge redevelopment challenges with new platforms and decommissioning of existing platforms. Furthermore infill drilling of more than 50 wells is premised and there exists IOR opportunities. All the redevelopment has to respect the constraints for oil and gas production and the rig availability. Results from the application is an optimised production strategy/drilling schedule for all fields in the area, leading to a significant increase in project value. The solution also resulted in a longer plateau production and higher ultimate reserves, through the drilling schedule optimisation. The proposed method offers a new method of LP formulation incorporating both influence functions, well scheduling effects and choking of wells, which allow iterative solutions. Introduction The most common method of field development evaluation is to investigate several basic plans which are based on experience and judgement. Based on development screening one concept is chosen for detailed study. In instances of combined redevelopment with a large number of drilling opportunities, multiple reservoir tie-in decisions, IOR evaluations, and processing/transportation limitations, a complete optimisation study is required. Initially the scope of such a study may appear prohibitive. This paper provides a formulation of optimisation and reservoir simulation which is manageable with a minimum use of time and resources. Linear Programming (LP) has previously been used to address situations as described above. Linear Programming is a term related to the systematic solution of a set of linear equations in order to optimise a given objective function. The formulation of a field development study into a LP problem presents many challenges, both with respect to linearisation of the problem and the size of the problem. In addition most real development cases will have to solved by Mixed Integer Programming (MIP), in order to allow predefined variables to assume the value of 0 or 1. These types of variables can be viewed as yes/no flags and are a key component of forming a development strategy. P. 497^

Abstract This paper describes numerical techniques developed to determine the wellbore configuration that will optimize the life-time value of a property. The production history of a reservoir can be predicted by integrating a reservoir model, a well-bore flow model, a choke model, and a separator model. Changes in any production parameter, including the gas-lift configuration, will cause changes in the predicted production history. The numerical methods can find the combination of production parameters that optimizes the net present value of the flowstream. The control parameters sampled in this work include tubing diameter, separator pressures, depth of gas injection, and volume of gas injected. Each of these parameters can be variable with time. The influence of these parameters upon the net present value is complicated by the feed-back nature of the gas-injection loop. This nonlinearity requires robust, efficient routines for optimization. Genetic Algorithm optimization techniques are shown to be both stable and efficient when used to optimize these sorts of nonlinear problems. Introduction Petroleum engineers face a wide variety of parameter estimation and optimization problems. Every time a transient test is analyzed or a location for a well is chosen, an optimization or parameter estimation problem has been solved. Yet it is surprising how rarely a formal optimization technique is used to solve these problems. Mathematical and heuristic methods are available which can find optimal solutions for a given model. This paper describes the application of formal optimization techniques to a realistic petroleum engineering problem. With the improvement of computer modeling over time, it is now feasible to apply optimization techniques to petroleum engineering models. In 1990, Carroll and Horne applied multivariate optimization techniques to a model of a field produced by a single well. However, only the separator model was compositional, and no wellbore parameters were allowed to vary with time. Carroll and Horne used two types of optimization routines, gradient methods and polytope methods. In 1992, Ravindran followed and allowed engineering parameters to be varied with time. Fujii continued this line of study in 1993, by including a network of wells connected at the surface. Fujii also studied the utility of genetic algorithms for petroleum engineering optimization. This project expands on Ravindran's work by replacing all black-oil components with fully compositional models. In addition, all three major types of optimization techniques were considered. Model Construction The model developed for this project was designed to test optimization algorithms for petroleum engineering problems. In addition to being robust, the model is a simplification of a real petroleum engineering problem. If the tested optimization algorithms are to be useful for real applications, the test problem should include challenging nonlinear relationships. The field model constructed for this work provides these challenges. The field model is an integration of smaller components. The complete model represents an oil reservoir with a single gas-lifted well. The individual components include: reservoir model; well model with gas lift; choke model; and separator model. Fig. 1 illustrates the fashion in which these components are combined. In the integrated model, each component influences the other components. Produced fluid flows into the tubing. At the point where the lift gas enters the tubing, the two streams commingle, and the flow continues up the wellbore. At the surface, the commingled fluid passes through the choke into the first separator. Gas from the first separator goes into the gas line and the liquid phase moves into the second separator. Gas from the second separator passes into the gas line, and the liquid goes into the third separator. The gas in the third separator goes into the gas line, and the liquid goes into stock tanks to be sold. Some of the separated gas is compressed, and injected into the tubing-casing annulus for gas lift. The remainder of the gas is sold. Notice the nonlinear relationships between the various components. P. 341^

Abstract As the reservoir pressure declines with time, production from a gas field drops. In order to meet delivery contracts, this invariably requires the addition of relatively expensive compression facilities that boost production by lowering the bottom-hole pressure and improving the drawdown across the sandface. This paper analyzes the compressor installation costs and operating strategy required to meet the production goals over the life of a gasfield, and proposes a simulation solution based on optimizing overall profitability. Key elements of the objective function to be optimized include a time-dependent revenue stream based on the projected price of gas, capital costs associated with adding incremental compression at periodic intervals, and compressor fuel consumption costs that are typically a near-linear function of the operating horsepower. The optimization model proposed is applied to an example gasfield for a fifteen-year operating strategy. The results demonstrate that an optimized compressor installation and operation strategy can improve overall field profitability considerably. The effect of varying gas price scenarios on the optimization solution is analyzed separately. Introduction The wells and gathering lines in a typical gas field are connected in a "branch-and-tree" type configuration, with single-well flowlines leading into common "branch" lines which in turn flow into the main line leading to separation, treatment and compression facilities. There is usually significant hydraulic interaction between the wells, with back-pressures at separation and treatment facilities controlling the overall rate of production. Increased production from a well results in a correspondingly higher back-pressures on adjacent wells that share a common flowline. This in turn causes a reduction in their rates of production. Single-well nodal analysis-based methods for evaluating productivity must therefore be eliminated in favor of a network model of the overall gathering system. As reservoir pressure declines, existing production from a field can only be maintained by adding compression. Long-term delivery contracts can require the operator to add compression several times over the life of the field. In this paper, we develop a time-dependent network model of a gas field. The model includes the interaction between wells, flowlines, surface facilities, and the reservoir. The interactions with the latter are accomplished by incorporating simple relationships between pressure decline and reservoir inflow. An economic model is then generated from the field model. This is accomplished by defining a profit function that measures the difference between gas revenues and the compression costs that drive the production of gas. Lastly, we maximize the profit function by considering variations in compressor operational strategies that are consistent with a long-term contract for the field. The optimization algorithm used is based on the Successive Quadratic Programming (SQP) technique, a nonlinear approach that has been applied successfully to problems with multiple decisions variables in the process simulation industry. The field-wide, time-dependent network model, coupled with the economic model of the compressor operation strategy, and integrated with the SQP optimization algorithm, has been developed with NETOPTTM, a commercial field-planning and optimization tool for single-phase and multiphase production systems. II. SIMULATION MODEL The field-wide simulation model is based on a generalized steady-state pressure-balance algorithm applicable to multiphase and single-phase production networks and pipeline systems. The network is defined in terms of sources (producing wells), sinks (delivery point), and junction nodes, that are connected by links. P. 511^

Abstract This paper describes a comprehensive approach to formation damage diagnosis. Once the cause of formation damage and the location of the damage is known, then one can select the optimum fluid and design the best treatment to remove the damage. Our approach combines the domain knowledge with the best available expertise, then uses new technologies, which include fuzzy logic and an expert system. Our approach first diagnoses the formation damage mechanism, then we determine the treatment method, and finally we select the optimal fluids for the treatment. Introduction The selection of treatment fluids for matrix stimulation is more complicated than for hydraulic fracturing. In matrix acidizing, we first have to determine the damage mechanism and the location of the damage; then, we can choose the proper treatment fluid based upon the damage type, lithology, bottomhole temperature, and formation permeability. Near wellbore formation damage can be caused by chemical or mechanical processes that occur in the formation very near the borehole. Formation damage resulting from drilling, completion, production, or stimulation treatments can substantially reduce the productivity or injectivity of a well. Different types of formation damage require different types of stimulation treatments to repair the damage and increase productivity. However. diagnosing formation damage is not a simple task. Diagnosing damage requires an adequate data set, careful analysis and experience. Multiple damage mechanisms may coexist, which makes the problem even more complicated. Many investigations have been conducted to learn how to diagnose formation damage. Numerous fluid systems and methods to remove formation damage have been developed. However, diagnosing the formation damage mechanism, then removing the damage from the formation or wellbore is a difficult problem. The problem becomes more difficult when adequate data are not available to properly define the problem. Two papers have been published that discuss how expert systems and rules can be used to assist in the analyses of formation damage problems; however, no complete approach to the problem of diagnosis, then treatment has been written. In this paper, we describe a complete approach to the problem. Applying an expert system and fuzzy logic techniques, we have develop a comprehensive approach and a computer model, that can be used to diagnose formation damage types. We also choose the optimal stimulation method and an optimal treatment fluid system to remove the damage. Our methods are especially useful when very little data are available and/or multiple damage mechanisms may coexist. The approach helps an engineer design the optimum matrix stimulation treatment. P. 811

Abstract Produced petroleum fluids containing hydrogen sulfide (H2S) Produced petroleum fluids containing hydrogen sulfide (H2S) gas are often referred to as "sour". Hydrogen sulfide gas in the presence of free water can cause certain metal alloys to crack which is referred to as Sulfide Stress Cracking (SSC). SSC of metals is a serious safety issue for the petroleum industry and is addressed by the NACE (National petroleum industry and is addressed by the NACE (National Association of Corrosion Engineers) MR-01-75 document which establishes requirements for metallic materials suitable for sour environments. Sour service materials selection is a complex problem often facing personnel not experts in this area. Since SSC and the NACE MR-01-75 document are not easily understood by most petroleum engineers, an expert advisor computer program was developed to integrate expert comments, SSC fundamentals, key metallurgy concepts, and NACE MR-01-75 requirements. This program is termed an expert advisor rather than an expert system because it seeks to teach the knowledge rather than simply provide expert answers. A novice user is able to proceed through a structured tutorial to establish the necessary fundamentals on metallurgy, SSC technology, and NACE MR-01-75 requirements necessary for informed use of the NACE document. More experienced users can proceed directly to the expert system portions of the system for direct evaluation of an alloy specification. A general survey option is available within the expert system portion to evaluate NACE requirements in a general context without being limited to the specifics of a given alloy specification. This software demonstrates that expert systems technology can be applied to SSC and NACE MR-01-75 to provide expert communication to users. Introduction The National Association of Corrosion Engineers NACE MR-01-75 document was developed as guideline to assist in the selection of alloys that are suitable for sour service as well as manufacturing requirements including heat treatment and cold working. The NACE MR-01-75 document is limited in scope. The focus of this document is on applications where safety is an issue if pressure containment is compromised. The NACE document assumes the user has a good understanding of the complex relationship between the environment and metallurgy. Petroleum engineers often do not have the necessary metallurgical background to make informed decisions about materials selection in cases where sour service is involved. An example is the relationship between heat treatment procedures, residual stresses, and SSC susceptibility. An expert system software package has been developed to bridge this "knowledge gap". This expert system differs from the traditional expert system since the approach has been to educate and to assist the user in making an informed decision. The software does not simply interrogate the user and supply a single answer. Novice users are able to learn metallurgical terminology and complex relationship between the environment, manufacturing requirements, and alloy selection. For the expert, the tutorial part of the program can be bypassed and the user is provided a systematic checklist in the materials selection process. P. 971

Abstract Determining an optimal fracture design can be a complex process, involving large amounts of computer and engineering process, involving large amounts of computer and engineering time. This paper describes a method that utilizes mixed integer-linear programming to adjust a treatment's fluid selection, pump rate, proppant volume, and frac dimensions to arrive at an optimal design. This method can be readily adapted to most any fracture design model. The optimization method presented in this paper has shown to be able to design less expensive fractures for a given stimulation ratio in less time than a trial and error approach. In the examples presented, variable stimulation costs (proppant, fluid, and presented, variable stimulation costs (proppant, fluid, and hydraulic horsepower charges) were reduced by 50% compared to trial-and-error-derived designs. Introduction This paper is divided into two portions to separate the applicability of this method from the rather dry linear programming formulation. The main section of the paper provides an overview of the method and an example application. The second section covers the detailed derivations in 2 appendices. Appendix 1 describes the linear program formulation. Appendix 2 describes the methodology used to convert the optimal frac selections into a treatment schedule. Linear Programming Overview. Linear programming (LP) is not a computer programming language, but rather a mathematical technique derived in the 1930's. The purpose of linear programming is to maximize or minimize some desired objective subject to certain constraints. For this paper', LP is used to determine the most economic combination of fluid type, fluid volume, pump rate, and proppant volume. This paper presents the formulation of the LP model, and includes presents the formulation of the LP model, and includes references for LP software. The Problem. Normally, a frac design is based on a combination of an engineer's experience and a number of design sensitivities. These sensitivities may evaluate relationships such as the cost I performance tradeoff of different frac fluids. The frac designer can find that while a high viscosity fluid may be very efficient, its high friction pressure may make it an expensive choice due to high hydraulic horsepower requirements. Additionally, it may be found that fracs of different dimensions give the same performance, but at widely different costs. The resulting design problem represents a 5 dimensional "Rubik's Cube" that an engineer must manipulate to determine the best design. The final-frac design may eventually be based more on relieving "calculation fatigue" than designing a cost-effective treatment. The final design will get the job done and certainly be profitable, but perhaps a treatment exists that will accomplish the same job with less cost. To do this, an LP technique is needed that will investigate all possible combinations of design parameters to determine the most cost-effective design. The Solution. The method presented in this paper does not replace any established frac design methods, but rather seeks to augment existing methods. To optimize a treatment, the following 3 steps are taken: A series of calculations are made to determine the performance of each fluid "candidate" at different pump rates. performance of each fluid "candidate" at different pump rates. P. 161

Abstract AL (Artificial Lift) is a software program for selecting the pumping processes best suited to an oil well. It integrates an expert system which manages the knowledge of the subject and organizes the implementation of a set of algorithmic programs for making technical assessments. AL was developed at the University of Missouri-Rolla, Petroleum Engineering Department. This article describes in detail the expert-system part and gives an example of the use of the complete system. Introduction Certain problems in the petroleum industry are being analyzed by a relatively new aspect of computer systems known as artificial intelligence (AI). To illustrate the use of expert system technology, a prototype rule based expert advisor program for selecting the optimal pumping method for an oil well has been developed. This program considers the four basic methods of pumping: Gas Lift, Electric Submersible Pump, Hydraulic Pump, and Sucker Rod Pump. The program starts by prompting the user for parameters related to the candidate well. As these questions are answered, AL narrows down the choices. Additional specific questions pertaining to the remaining lift methods are pertaining to the remaining lift methods are asked as AL continues to narrow down the choices. Finally AL gives its EXPERT opinion of the optimal lift method. The user, whether an undergraduate petroleum engineering student or an experienced engineer, is reminded of the factors that influence the selection of the artificial lift method while answering the questions. Background Expert systems are computer programs that use expert knowledge to attain high levels of performance in a narrow problem area. The individual human expert skills needed to interpret and assess data (decide which lift method is best in a given set of circumstances) are a result of years of formal training, rules of thumb, methods of dealing with constraints and human prejudice. In short humans have gained through experience knowledge. AL differs from conventional programs in that knowledge is represented explicitly, as well as, data and algorithms, and that knowledge can be easily enhanced through increased exposure to a particular problem area. AL has a knowledge base, (expertise is represented in a series of IF-THEN rules), an inference engine (processes rules according to some control strategy), and a database (where facts, deductions and intermediate answers are stored). The control strategy could be data driven or goal driven, or both. In a data driven strategy, the rules are worked forward checking against the database to make further conclusions. In AL, a goal driven strategy, the system starts with a hypothesis and by using rules that establish the hypothesis, works backwards, checking their preconditions against that database. As and expert systems AL has the following characteristics: Symbolic representation of knowledge; if temperature greater than TEMPERATURE and temperature greater than 250 then centrifugal lift is unsuccessful. P. 423