IBP1342_06

EVALUATION OF NEURO-SIMULATION TECHNIQUES AS

PROXIES TO RESERVOIR SIMULATOR
Paulo Camargo Silva1, Célio Maschio2, Denis J. Schiozer 3

Copyright 2006, Instituto Brasileiro de Petróleo e Gás - IBP

Este Trabalho Técnico foi preparado para apresentação na Rio Oil & Gas Expo and Conference 2006, realizada no período de 11 a
14 de setembro de 2006, no Rio de Janeiro. Este Trabalho Técnico foi selecionado para apresentação pelo Comitê Técnico do evento,
seguindo as informações contidas na sinopse submetida pelo(s) autor(es). O conteúdo do Trabalho Técnico, como apresentado, não
foi revisado pelo IBP. Os organizadores não irão traduzir ou corrigir os textos recebidos. O material conforme, apresentado, não
necessariamente reflete as opiniões do Instituto Brasileiro de Petróleo e Gás, seus Associados e Representantes. É de conhecimento e
aprovação do(s) autor(es) que este Trabalho Técnico seja publicado nos Anais da Rio Oil & Gas Expo and Conference 2006.

Abstract
Proxy models are becoming more widely used as they can simplify highly complex processes with reasonable accuracy.
Process such as production history matching, well placement, risk analysis, for example, generally involve large number
of reservoir simulations and large computational effort. To minimize this problem some techniques such as Spline,
Neural Networks, Kriging and Experimental Design, have been presented in the literature to be used as proxies to
reservoir simulator. Due to the importance of the decisions related to the development and management of petroleum
fields, the development of proxies with high accuracy can be a decisive aspect in a project. The successful applications
of Neural Networks in several research fields suggest the investigation of appropriated architectures to be used as
proxies to reservoir simulator. In this article novel proxies to reservoir simulator, based in Neuro-Simulation
Techniques, are presented. These proxies present a high accuracy to reservoir simulator. Five different architectures of
Neural Networks were studied and applied in two case studies related to history matching. The results obtained showed
the large potential of application of the techniques introduced.

______________________________
1
Dsc Civil Engineering, Researcher at UNISIM - UNICAMP/FEM/DEP
2
Dsc, Mechanical Engineer, Researcher at UNISIM - UNICAMP/FEM/DEP
3
PHD, Petroleum Engineering - Professor at UNICAMP/FEM/DEP
 Rio Oil & Gas Expo and Conference 2006

1. Introduction
In the last years is growing the development of researches related to generation of proxies to reservoir
simulators. Neural network based proxies, for example, is being very used in several process involving reservoir
simulation (Singh, 2005; Reis, 2006; Zangl, 2006, Al-Thuwaini, 2006). The common methodologies generally used for
the elaboration of the proxies are: thin plane spline; kriging; experimental design and artificial neural networks. There
are two principal criteria for evaluation of proxies (1) The proxies must exactly reproduce the data which they used in
the their elaboration; (2) The correlation between the data produced by the proxies and the real data it is satisfactory if
the value is between 0.7 and 0.8. In this article are presented the results of researches related to application of several
different types of Artificial Neural Networks as proxies to reservoir simulators. It is showed that determined types of
Artificial Neural Networks are able to yield proxies of high accuracy. Two case studies related to History Matching are
examined, a synthetic case and a real case. These cases were used only as data source for the proxies generation. The
proxies were not effectively applied in the history matching process.

2. Methodology
The methodology developed in this research presents the following main steps:
(1) To create knowledge base with reservoir simulations;
(2) To train, to test and validate Artificial Neural Networks with the reservoir simulations of the knowledge base;
(3) To apply the proxy in the simulation parameters;
(4) To compare the results of the proxy with results of the reservoir simulator.
In this methodology, reservoir simulations compose a knowledge base. The simulations of this knowledge base are used
in the training, test and validation of the Artificial Neural Networks. When the Neural Networks are trained, tested and
validated these networks constitute proxies. These proxies are applied in parameters of reservoir simulations which were
not used to create the proxies and the results are compared with the results of the reservoir simulator. Along with the
work are showed comparisons between results obtained from proxies and from the reservoir simulator in two case
studies related to History Matching. In order to generate the proxies, several Artificial Neural Networks were used:
(1) Radial Basis Network (RBN) - The RBN is a known Neural Network, often applied to approximate
functions, Orr [1995]. The basic architecture of the RBN which we implemented consists in a network with three layers:
(1) The first layer is an input layer; (2) The second layer is an hidden radial basis layer, with transfer function e-n2,
where n is the distance between the input vector and weights of the links, multiplied by a bias; (3) The third layer is an
output layer with a linear function. Two types of this basic architecture were implemented:
Type 1: With a fixed number of neurons in the hidden radial basis layer. This number is equal to the number of samples
in the training set;
Type 2: With the neurons in the hidden radial basis layer added at each interaction until the minimal sum-squared error.
(2) Adaptive Neuro Fuzzy Inference System (ANFIS) - This fuzzy inference system was applied for the data
modeling. From a set of Input/Output data related to a multiple input simple output system a fuzzy inference system was
produced. The parameters of this system were estimated with base in Neuro-adaptive learning techniques. These
techniques operate as in the conventional learning of the Neural Networks, but they use a combination of the least
squared estimation and back-propagation algorithms, Jang [1993].
(3) Fuzzy System with Subtractive Clustering (FSSC) - Given a set of Input/Output data of a multiple input
simple output system, a cluster analysis was realized by using the technique of subtractive clustering Chiu [1994]. The
centers of the clusters were used to define the rules of a fuzzy inference system. That is, a fuzzy inference system was
structured with base in rules extracted by using subtractive clustering of a set of Input/Output data.
(4) Generalized Regression Neural Network (GRNN) - This Neural Network has been used in many
applications to approximate functions. We implement the following architecture of this Neural Network: A network with
three layers: (1) The first layer is an input layer; (2) The second layer is an hidden radial basis layer with the number
neurons equal to the number of samples in the training set; (3) The third layer an output layer. The output layer of this
network has a linear function, but the input of this linear function is the output of the second layer multiplied by weights
submitted to a normalization function Shoiad et.al. [2004].

3. Case Study – History Matching

3.1. Case1
In this case, a synthetic reservoir derived from SPE 10 was used. The characteristics of the reservoir for this
case are described in Maschio et al. [2005] and in Maschio and Schiozer [2004]. The reference model was used to
generate a synthetic history and the coarse model is to be adjusted constrained to the history generated through the fine
grid. The matching parameters used were the horizontal permeability of the six layers, with 27 multipliers between 0.3
2
 Rio Oil & Gas Expo and Conference 2006

and 2.5. The objective function (difference between simulated and observed data) was composed by the water rate of the
4 producer wells.

3.2. Case 2
Case 2 is derived from a real reservoir. The main characteristics of the case is also describe in Maschio et al.
[2005] and Maschio and Schiozer [2004]. The objective function was composed by field pressure, field water rate and
field oil rate. The matching parameters (all of them in the field domain) was porosity and net to gross, both with 9
multipliers ranging between 0.7 and 1.3, horizontal permeability with 9 multipliers between 0.4 and vertical
permeability with 9 multipliers between 0.3 and 2.5.

4. Comparisons of the proxies with the reservoir simulator

In this section are showed graphical representations of the comparisons between the proxies and the reservoir
simulator. As mentioned above, the two cases studies are related to History Matching. So that objective functions (FO)
were obtained from reservoir simulations to adjust parameters of the reservoir model. The associations between
parameters and the values of the objective function to train, to test and validate the proxies are used. When the proxies
were trained, tested and validated are compared the results of the objective functions obtained from proxies with the
results of the objective functions obtained from reservoir simulator. Figures 1, 2, 3 and 4 are related to Case 1. Figures
5, 6, 7 and 8 are related to Case 2. In these Figures the blue curves represent the responses of the reservoir simulator and
the green curves represent the responses of the proxies. Note that the proxies present high accuracy.

Simulator
Proxy

Simulations
Figure 1 - Adaptive Neuro Fuzzy Inference System (ANFIS), Case 1

3
 Rio Oil & Gas Expo and Conference 2006

Simulator
Proxy

Simulations
Figure 2 - Radial Basis Network (RBN) - Type 1, Case 1

Simulator
Proxy

Simulations

Figure 3. Radial Basis Network (RBN) - Type 2, Case 1

4
 Rio Oil & Gas Expo and Conference 2006

Simulator
Proxy

Simulations
Figure 4 - Fuzzy System with Subtractive Clustering (FSSC), Case 1

Simulator
Proxy

Simulations
Figure 5. Fuzzy System with Subtractive Clustering (FSSC) - Case 2

5
 Rio Oil & Gas Expo and Conference 2006

Simulator
Proxy

Simulations
Figure 6. Radial Basis Network (RBN) - Type 2, Case 2

Simulator
Proxy

Simulations
Figure 7. Generalized Regression Neural Network (GRNN), Case 2

6
 Rio Oil & Gas Expo and Conference 2006

Simulator
Proxy

Simulations
Figure 8. Radial Basis Network (RBN) - Type 1, Case 2

The number of reservoir simulations used for the elaboration of the proxies in all the cases was inferior to 50 %
of the available simulations. This meant a considerable reduction of simulations. The proxies elaborated can be applied
in studies that require a very large number of reservoir simulations, for example optimization studies. In these studies the
proxies can work directly with optimization methods, for example genetic algorithm, with considerable reduction of the
reservoir simulations.

5. Conclusion
In this article are showed that proxies to reservoir simulators with high accuracy can be obtained from Artificial
Neural Networks. Different neural networks were examined in two case studies of history matching: a synthetic case and
a real case. The proxies showed good agreement to the solution obtained from simulator, which permits to conclude that
for cases tested, it could be used in the optimization process. The use of neural networks is an promise area in the
processes involving reservoir simulations.

6. Acknowledgement
The authors would like to thank the Conselho Nacional de Desenvolvimento Científico e Tecnológico
(CNPq/PROSET), FINEP/CTPETRO, PRH/ANP, and PETROBRAS for their financial support.

7. Referências
AL-THUWAINI, J.S. ZANGL G. PHELPS R. Innovative Approach to Assist History Matching Using Artificial
Intelligence, SPE 99882, Intelligent Energy Conference and Exhibition, 11-13 April, Amsterdam, The Netherlands,
2006.
MASCHIO, C.; SCHIOZER, D. J AND MOURA FILHO, MARCOS A. B. A Methodology to Quantify the Impact of
Uncertainties in the History Matching Process and in the Production Forecast SPE 96613, SPE Annual Technical
Conference and Exhibition, 9-12 October, Dallas, Texas, U.S.A, 2005.

7
 Rio Oil & Gas Expo and Conference 2006

MASCHIO, C. AND SCHIOZER, D. J. “Ajuste de Histórico Assistido Usando Métodos de Otimização de Busca
Direta”. IBP 06204, Anais do Rio Oil & Gas Conference, Rio de Janeiro, 4-7 Outubro, 2004.
ORR, M. Regularization in the Selection of Radial Basis Function Centers, Neural Computation, 7(3), pp.606-623,
1995.
JANG, J.S. ANFIS: Adaptive Neural Network based Fuzzy Inference System”, IEEE Trans. On System, Man and
Cybernetics, Vol.23, No.23, May/June, pp. 665-685.
REIS, L. C. Risk Analysis With History Matching Using Experimental Design or Artificial Neural Networks SPE
100255, SPE Europec/EAGE Annual Conference and Exhibition, 12-15 June, Vienna, Austria, 2006.
SINGH, S. Permeability Prediction Using Artificial Neural Network (ANN): A Case Study of Uinta Basin, SPE 99286,
SPE Annual Technical Conference and Exhibition, 9-12 October, Dallas, Texas, U.S.A, 2005.
SHOIAD, M.; AWAIS, M.; MASUD, S.; SHAMAIL, S.; AND AKHTAR, J. ‘Application of Concurrent Generalized
Regression Neural Network for Arabic Speech Recognition”. Proceedings of Neural Networks and Computation
Intelligence, NCI 2004.
ZANGL, G. AND GRAF, T. Proxy Modeling in Production Optimization, SPE 100131, SPE Europec/EAGE Annual
Conference and Exhibition, 12-15 June, Vienna, Austria, 2006.