RISK ANALYSIS APPLIED TO PETROLEUM FIELD DEVELOPMENT:
AUTOMATED METHODOLOGY AND PARALLEL
COMPUTING TO SPEEDUP THE PROCESS
E.L. Ligero and D.J. Schiozer Departamento de Engenharia de Petrleo, Faculdade de Engenharia Mecnica UNICAMP. P.O.Box 6122, Zip Code 13081-970, Campinas, So Paulo, Brasil. (eligero@dep.fem.unicamp.br)
1. INTRODUCTION
The petroleum field life cycle has the following stages: exploration, appraisal, development, production and abandonment. The appraisal phase is strongly related to uncertainties, high investment and field development decisions. There are three main types of decisions involved in this phase: abandonment of the discovered field (low profit, low stock tank oil originally in place (STOIP), high oil viscosity, etc.), continuation of the appraisal phase (risk mitigation) and development of the field. A petroleum field development requires large investments and any improvement in the process can represent significant additional profit. However, in the 80s, it was usual production forecast based in a deterministic reservoir simulation model. In this way, production forecast had a deterministic approach. In most of cases, production forecast was optimistic. Thus, economical viability of the project of oil and gas fields was guaranteed by high prices practiced in the market. The current tendency is production forecast obtained by a probabilistic approach. This approach is based in numeric flow simulation of several models representing uncertainties of a petroleum reservoir and allows the evaluation of the uncertainty performance, like cumulative productions and oil rates, in any simulated time. Important advances in hardware development allow increasing accuracy in production prediction in the decision processes. Furthermore, for complex reservoirs and large fields, probabilistic approach is possible and highly necessary in the production strategy definition. The application of a probabilistic approach requires the definition of a methodology, which must have the following characteristics: easy to use, flexible, and applicable to a large range of cases. However, special attention must be given to reduce excessive computation effort and to minimize the total time of the process. In order to achieve this, it is necessary to make some simplifications to improve performance without losses in precision. The objective of this paper is to show some advantages of having an automated methodology to perform risk analysis in Exploration and Production Project, during the appraisal phase, using reservoir simulation. Parallel computing can be used to reduce total processing time. As an example, the methodology is applied to a synthetic reservoir model. During the appraisal phase, the use of reservoir simulation is necessary to obtain more accurate results of several economic and technical parameters. However, the use of numerical simulation increases substantially the computational effort. Therefore, automated procedures are necessary to make the process viable.
Base Model Attribute1-P Attribute1-M Attribute1-O Attribute2-M Attribute2-P Attribute2-O Attribute2-M Attribute2-P Attribute2-O Attribute2-M Attribute2-P Attribute2-O Prob(1P_2P) Prob(1P_2M) Prob(1P_2O)
Prob(1M_2P) Prob(1M_2M) Prob(1M_2O) Prob(1M_2P)
Prob(1M_2M) Prob(1M_2O) 2. METHODOLOGY
The risk analysis methodology used here was first developed by Loschiavo et al. (2000) and implemented by Steagall and Schiozer (2001). It is based on the simulation of several flow models that represent the possible scenarios of the reservoir, through the combination of the uncertain attributes. The simulation models are built through the derivative tree, as in Figure 1. Each final branch of the derivative tree corresponds to a complete simulation model. The probability of each resulting model is equivalent to the multiplication of the probability of attributes that composes it. The sum of probabilities for the models represented by branches in the tree must be equal to 1. For a simple example as illustrated in Figure 1, the total number of models to be simulated is equal to 9 (3 2 ). The inclusion of two more attributes, with three uncertain levels each one, it would increase the number of models for 81 simulations. To reduce the number of models to be simulated, a sensitivity analysis is applied to uncertain attributes in order to select the critical ones. This is one important simplification that reduces the total time for the risk analysis process. Examples of usual uncertain attributes are the structural model, horizontal and vertical permeability, relative permeability, pore volume, rock compressibility, PVT curve, etc. The sensitivity analysis is done changing the attributes one by one in the base model. The base model is equivalent to the deterministic model and is generated by combination of medium or probable values for all attributes. This analysis can be made in one or more simulation times considering different production or economical parameters such as Cumulative Oil, Gas and Water Production or Net Present Value (NPV).
Figure 1: Example of Derivative Tree with Two Attributes and Three Levels (P - Pessimistic, M Medium and O - Optimistic) (Schiozer et al., 2002)
The generated models are simulated and the results are compared each one with the base model. Thus, it is possible to choose the critical attributes. Once critical attributes are chosen, it is possible to elaborate the derivative tree. It is important to have in mind that only the branches, which are constituted by critical attributes, are submitted to simulation runs.
3. AUTOMATED PROCESS AND PARALLEL COMPUTING
In automated process, reservoir simulation models covering the sensitivity analysis are automatically generated and sent to a commercial Black-Oil simulator. These models are submitted to flow simulation with parallel computation in a network of workstations. After all simulation processes during sensitivity analysis, results are stored in order to avoid their unnecessary simulation later. The sensitivity analysis is obtained in a graphic interface where the critical attributes can be selected. Flow models, resulting from the combination of the critical attributes, are built automatically through the derivative tree technique. An important aspect is that only models that have not been simulated in the sensitivity analysis are submitted to the flow simulation. After simulation of the derivative tree models, a statistic treatment is done to obtain a risk curve, or an expectation curve, of the production forecast. The final results are individual results of each model and cumulated percentiles of all simulated models. The percentiles, usually, calculated are P 90 , pessimistic, P 50 medium or most probable and P 10 , optimistic. This automated procedure and parallel computing used here are illustrated in Figure 2. Building process of simulation models, simulation runs and statistical analysis of the results are executed automatically, thus providing a great reduction in time.
Reservoir Parameter Uncertainty Studies Sensibility Analysis Flow Models Decision Tree Building (uncertainty representatives) Parallel Computing Simulation Statistical Treatment Uncertainty Production Forecasts P10, P50, P90, Models Risk Curves Reservoir Parameter Uncertainty Studies Sensibility Analysis Flow Models Decision Tree Building (uncertainty representatives) Parallel Computing Simulation Statistical Treatment Uncertainty Production Forecasts P10, P50, P90, Models Risk Curves
Figure 2: Automated Process of Risk Analysis in Exploration and Production Projects (Steagall and Schiozer, 2001)
The automated process must be flexible. Some possible alternatives that dont require additional simulations are: changing the occurrence probability of critical attributes and modifying the economic data model. These modifications can be done very fast, since they dont change the simulation results, which are stored to avoid unnecessary simulation runs. It is highly recommendable to execute a sensitivity analysis regarding uncertainty in economical parameters, which permit having a range of possible economical scenario with no extra simulations. The application of the derivative tree methodology may increase the computational effort, but in general, the time required for the whole process decreases if the computational tool is used in a correctly. Parallel computing is responsible for the speedup of the process. Though this speedup can be observed, it is difficult to be measured. Figure 3 shows the speedup caused by parallel computing. The process can be executed 5 to 10 times faster, if a typical network of 10 to 15 workstations is correctly implemented.
Figure 3: Speedup Measured and Trend for Homogeneous and Heterogeneous Network. (Schiozer et al., 2002)
4. APPLICATION
4.1. Description
The global reservoir has 20x44x17 cells (14,496 cells) and it was obtained through an upscaling procedure from the Model 2 presented by Christie and Blunt (2001). The coarse grid is described on a regular Cartesian grid and its dimensions are 183 x 152 x 3 m. A fixed five-spot production strategy was adopted. An important reservoir characteristic is the absence of gas, in other words, the fluids contained in the reservoir are only oil and water. Considered uncertain attributes were structural model, porosity, horizontal permeability, vertical permeability, relative permeability and water-oil contact. These uncertainties and their respective values and occurrence probabilities are illustrated on Table 1. Since the structural model is an uncertain attribute, area0 denotes the global structural model (20x44x17) and area1 represents another possible model with 15x44x17 (11,220 cells). Excluding structural model and water-oil contact, that have 2 levels of uncertainty, all attributes have three levels of uncertainty, a probable or a medium (level 0), an optimistic and another pessimistic. If all attributes on Table 1 were combined, the total number of possible reservoir simulation model would be equal to 324 (2 2 * 3 4 ). The simulation of this high model number is sometimes not feasible. The reduction of the number of simulations is possible through the sensit ivity analysis, whose main purpose is to exclude the attributes with little influence on a specific function.
In most cases of derivative tree application, such as in the studied case, critical attributes are not known at the beginning. In the sensitivity analysis it is necessary to choose one or more functions and simulation times in order to determinate the critical attributes. The main functions to be chosen are: Cumulative Oil Production (Np), Cumulative Gas Production (Gp), Cumulative Water Production (Wp), Net Present Value (NPV), Internal Rate of Return (IRR) and Return over Investment (NPV/Investment). Three functions, in the studied case, were selected to illustrate the sensitivity analysis procedure. Two of them are directly related to production, Cumulative Oil Production and Cumulative Water Production and the third one is an economical function, Net Present Value. On the other hand, only two simulation times were specified in order to calculate the values of these functions, 3650 days (10 years) and 7300 days (20 years). The number of reservoir simulation models generated in sensitivity analysis is, in this case, 11, comprising 1 corresponding to base model, 8 models (4 attributes with 2 extra levels each) and 2 models (2 attributes with 1 extra level each). The Figure 4 shows the sensitivity analysis results. Figures 4(a) to (f) show that sensitivity analysis depends on the function (VPL, Np or Wp) and simulation time (10 or 20 years).
(a) (b)
(c)
(d)
(e) (f)
Figure 4: Influence of Function and Time in Sensit ivity Analysis (a),(b) Net Present Value, (c),(d) Cumulative Oleo Production, and (e),(f) Cumulative Water Production. The sensitivity analysis by itself is a simplification and two important decisions can be taken after it: reduction of the number of uncertain attributes and reduction of number of levels of each critical attribute. A very important characteristic of the automated procedure is the addition, progressively, of critical attributes and/or levels until there are no more significant modifications in the values for analyzed function. It is important to emphasize if levels are added or subtracted, the probabilities of each level must be recalculated based on the characteristic of attribute.
4.3. Critical Attributes and Risk Curve
Once critical attributes are chosen, it is possible to build the derivative tree and estimate the production risk. The automated methodology is capable of building the derivative tree from the specified critical attributes. The models were simulated in sensitivity analysis are not simulated anymore. In this way, it is possible to reduce the required time to simulate all branches of the tree. In order to illustrate the results obtained from the statistical treatment of simulation models that constitute derivative tree, only one function and one time were chosen from Figure 4. Then the sensitivity analysis and risk were made for Net Present Value (NPV) for 20 years of production, Figure 4(b). The influence of critical attribute number was evaluated considering two groups of attributes: Group 1 with four attributes, dwoc (level 0 and 1), area (level 0 and 1), por (level 0, 1 and 2) and kr (level 0 and 1) and Group 2 with five attributes, dwoc, area, por, kr (level 0 and 1) and kz (level 0, 1 and 2). In both groups, kr had one uncertain level excluded by the sensitivity analysis then its probabilities were recalculated to 50% to Kr0 and 50% to Kr1. Figure 5 shows the percentiles P 10 , P 50 and P 90 considering the two groups of critical attributes. It is possible to observe that the base case is very close to the optimistic model (P 10 ). A very small difference between the estimated P 10 , P 50 e P 90 to four and five critical attributes is shown. The consideration of five critical attributes increases the simulation number from 18 to 46. It was unnecessary to consider five critical attributes, since only four could represent the reservoir uncertainty with almost the same precision. Therefore, it would be not necessary to add kz as critical attribute, since additional simulations could be avoided. The uncertainty is quantified by risk curve. The risk curve depends on the number of attributes and their values, the number of levels and their associated probabilities. The comparison between risk curves considering four and five critical attributes are considered. The estimated Net Present Value for 20 years, considering the defined critical attributes on Group 1 and 2, is on Figure 6. The risk of Cumulative Oil Production forecast for 20 years, with the same critical attributes like in NPV, is illustrated on Figure 7. In both figures, Net Present Value and Cumulative Oil Production risk curves are practically coincident and they are in concordance with the results on Figure 5. The results of the simulated models that constitute the derivative tree can be plotted as shown in Figures 8 and 9, which represent the typical uncertainty in Cumulative Oil Production and Cumulative Water Production, respectively.
Figure 5: Net Present Value for 20 years 4 and 5 Attributes
Figure 6: Net Present Value Risk Curve
Figure 7: Risk Curve of Cumulative Oil Production
-20 0 20 40 60 80 100 P 1 0 P 5 0 P 9 0 B A S E N P V
( M M
U S $ ) 4 Attributes 5 Attributes 0 10 20 30 40 50 60 70 80 90 100 -25 0 25 50 75 100 125 150 NPV (MM US$) C u m u l a t i v e
P r o b a b i l i t y
( % ) 4 Attributes 5 Attributes 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 Np (MM m3) C u m u l a t i v e
P r o b a b i l i t y
( % ) 4Attributes 5 Attributes
Figure 8: Uncertainty in Cumulative Oil Production
Figure 9: Uncertainty in Cumulative Water Production.
4.4. Discussion
The most common simplification in the automated risk analysis process is the sensitivity analysis, which is directly related to the critical attribute number and their levels. In addition to the sensibility analysis, other simplifications can be performed in order to decrease the computational effort. Examples of these simplifications can be seen in Ligero and Schiozer (2001) who used upscalled grid with a fewer number of 0 10000000 20000000 30000000 40000000 50000000 60000000 70000000 80000000 0 730 1460 2190 2920 3650 4380 5110 5840 6570 7300 Time (days) N p
( m 3 ) Structural Model 1 (area0) Structural Model 2 (area1) 0 10000000 20000000 30000000 40000000 50000000 60000000 70000000 80000000 90000000 0 730 1460 2190 2920 3650 4380 5110 5840 6570 7300 Time (days) W p
( m 3 ) Structural Model 1 (area0) Structural Model 2 (area1) blocks and in Costa and Schiozer (2002) who considered fewer levels of uncertainties for attributes that were considered not critical and aggregated several attributes in the sensitivity analysis. Other typical and usual simplifications are the use of production strategies as function only of the structural models, aggregation of variables, etc. In such cases, representative models (Steagall and Schozer, 2001) can also be very useful in the integration of geological and economical uncertainties. Representative models for the studied case are shown in Figure 10, which presents the relationships between Net Present Value and Recovery Factor of Oil of the models with uncertainties. Representative P 10 , P 50 and P 90 models are indicated through circles. It is possible to observe two groups of models, each group representing different values of dwoc. The distribution of points of this example if different of all other examples tested here because of the low number of simulations and high sensitivity of the water-oil contact These representative models can be used for several applications, for instance, to perform an economic sensitivity analysis. It is not necessary to use all simulation models to make this analysis. In addition, changing only economic models do not require extra simulations, since reservoir parameters do not depend on economic factors. The NPV variation caused by changing oil price is on Table 2. Santos and Schiozer (2003) have shown that different production strategies can be incorporated in the procedure in an automated way. Schiozer, Maschio and Ligero (2003) and Christie et al. (2002) used streamline simulation to reduce the computational time of each simulation run performed for the prediction the uncertainties in petroleum reservoirs. Many other simplifications are possible but the procedures will not be used if they are not automated because of the excessive amount of work involved.
Figure 10: Representative Models P 10 , P 50 and P 90 of NPV
-20000000 0 20000000 40000000 60000000 80000000 100000000 120000000 10 15 20 25 30 35 40 45 50 Recovery Factor of Oil (%) N P V
( U S $ ) Structural Model 1 (area0) Structural Model 2 (area1) dwoc0 dwoc1 P 10
In the 80s, production forecast used to be performed deterministically. Nowadays, the use of a suitable methodology, such as derivative tree, associated with numerical simulation flow, allows probabilistic predictions. The presented automated methodology to quantify exploration and production risk can be easily applied to most type of petroleum reservoir, from simple to complex models. It has the advantage to store simulation results, avoiding unnecessary simulation runs. In addition, parallel computing speedup the process, which may be fast until to complex reservoirs. In order to simplify the process, sensitivity analysis must be performed, which reduces the number and/or level of uncertain attributes. So the derivative tree is built, only, with critical attributes, reducing the model number and the global process time. Normally, the ideal number of critical attributes varies from 4 to 7, and they are included in the tree one by one. The process is interrupted when the required precision is reached. If critical attributes are more than 7, other simplifications are necessary to reduce the computational time and computational effort. Automated process permits to determinate representative models, allowing optimizing reservoir development plan for probable, optimistic and pessimist models with different geological characteristics.
6. REREFENCES
Christie, M., Subbey, S., Sambridge, M., Thiele, M., 2002. Quantifying Prediction Uncertainty in Reservoir Modelling Using Streamline Simulation, 15th ASCE Engineering Mechanics Conference, June, New York, USA
Christie, M.A. and Blunt, M.J., 2001, Tenth SPE Comparative Solution Project: A Comparison of Upscaling Techniques, SPE Reservoir Simulation Symposium, February, Texas, USA, SPE 66599
Costa, A.P.A. and Schiozer, D.J., 2002, Escolha de Atributos na Anlise de Risco em Campos de Petrleo na Fase de Desenvolvimento, 9th Brazilian Congress of Thermal Engineering and Sciences (ENCIT), October, Caxambu, Brasil.
Ligero, E.L. and Schiozer, D.J., 2001, Importncia da Escala do Modelo de Simulao de Reservatrios de Petrleo na Anlise de Incertezas em Projetos de Explorao e Produo, I Congresso Brasileiro de P&D em Petrleo e Gs, November, Natal, Brasil.
Loschiavo, R., Steagal D., and Schiozer D.J., 2000. Estudo do Impacto de Incertezas no Desempenho de Reservatrios de Petrleo, 8th Brazilian Congress of Thermal Engineering and Sciences (ENCIT), Porto Alegre, Brasil.
Santos, J.A.M., Schiozer, D.J., 2003, Quantifying Production Strategy Impact in Risk Analysis of an E&P Project Using Reservoir Simulation, 17th Reservoir Simulation Symposium, February, Houston, USA (to be published).
Schiozer, D.J., Maschio, C., Ligero, E.L. 2003, Quantifying the Impact of Grid size, Upscaling and Streamline Simulation in the Risk Analysis Applied to Petroleum Field Development , 17th Reservoir Simulation Symposium, February, Houston, USA (to be published).
Schiozer, D.J., Ligero, E.L., Costa, A. P.A., Santos, J.A.M, 2002. Risk Assessment for Reservoir Development under Uncertainty, Journal of Petroleum Science and Engineering (to be published).
Steagall, D.E. and Schiozer, D.J., 2001. Uncertainty Analysis in Reservoir Production Forecast during the Appraisal and Pilot Production Phases, SPE Reservoir Simulation Symposium, February, Dallas, USA, SPE 66399.