A Modular Signal Processing Model for Permeability Prediction in Petroleum Reservoir

The use of Artificial Neural Network (ANN) especially Backpropagation Neural Network (BPNN) has been a promising tool for well log analysis in predicting permeability. However, due to the range of permeability data, it is normally converted using logarithmic transform before being used for data analysis by the BPNN. This has an impact on the accuracy of the permeability prediction. This paper suggests a model for improving the permeability prediction. It first divides the whole sample space of the permeability values according to their logarithmic region, and then generates individual BPNNs for each logarithmic region. In this initial study, Learning Vector Quantisation (LVQ) is used for this purpose for separating the data. After that, each region is then handled by each BPNN. This method not only preserves the resolution of the permeability, but at the same time, increase the prediction accuracy. The contributions of this paper are to identify the problems in the signal processing of permeability prediction, and exploit new direction of improving permeability prediction using well logs. INTRODUCTION One of the key issues in reservoir evaluation using well logs is the prediction of petrophysical properties such as porosity and permeability. Of these petrophysical properties, permeability is one of the more important properties. Over the life of the reservoir, many crucial decisions depend on estimates of formation permeability. Permeability is widely used to determine the well production rate of the hydrocarbon, such as oil or gas. It is used to measure the fluid mobility that flows through the porus media when a pressure gradient is applied. However, the prediction of such properties is complex as the measurement sites available are limited to isolated well locations. Normally boreholes are drilled at different locations around the region. Well logging instruments are then lowered into the borehole to collect data at different depths known as well log data. Well logging instruments used in the measurement 0-7803-6278-0/00$10.00 (C) 2000 IEEE 906 Authorized licensed use limited to: Murdoch University. Downloaded on October 30, 2009 at 03:11 from IEEE Xplore. Restrictions apply. of well log data broadly fall into three categories: electrical, nuclear and acoustic [l]. Examples are Gamma Ray (GR), Resistivity (RT), Spontaneous Potential (SP), Neutron Density (NPHI) and Sonic interval transit time (DT). Beside the well log data, samples from various depths are also obtained and undergo extensive laboratory analysis. This laboratory analysis data is known as core data in the well log analysis process. In well log analysis, the objective is to establish an accurate interpretation model for the prediction of petrophysical properties for uncored depths and boreholes around that region [2]. There are three main widely used approaches for permeability prediction; namely empirical, statistical and ANN [3]. Recently, the use of fuzzy system [4] and fuzzy neural networks [ 5 ] have also emerged. Although the methods used are different, their objective is similar. It is mainly to establish an interpretation model by ways of linear or non-linear curve fitting. The ways they handle the processing of permeability data is also quite similar regardless of the method used. The next section of this paper will examine some of the possible problems in handling the permeability data. Section three will present a model that will improve the permeability prediction. Section four will examine the possible use of Learning Vector Quantisation (LVQ) in establishing the model discussed. Results and discussion of the test cases are also presented in this paper to show the proposed modular signal processing model could improved the accuracy of the permeability prediction. PROBLEMS OF PREDICTING PERMEABlLlTY Among most methods used in permeability prediction, ANN especially BPNN seems to be the most promising one in the literature [6],[7]. BPNN is the most popular among all ANN techniques in permeability prediction mainly because it is quite similar to Multiple Regression [8]. The analysis of the problems presented in this section will be based on the BPNN approach. However, most problems discussed here are also valid in other approaches used in permeability prediction. The problems mentioned in this section are discussed without taking any geology and petrophysics theory into consideration as these have been investigated in the geophysics literature. The analysis presented here is mainly viewed from the perspective of signal processing. The problems can mainly be divided into the following three areas:1. The normalisation of permeability values In most cases for ideal operation, BPNNs should only take values between 0 and 1 as input. Permeability values have to be normalised before they can be used in BPNN. There are normally two ways of performing normalisation in permeability prediction; they are linear or logarithmic transform. 0-7803-6278-0/00$10.00 (C) 2000 IEEE 907 Authorized licensed use limited to: Murdoch University. Downloaded on October 30, 2009 at 03:11 from IEEE Xplore. Restrictions apply. The equation that is used for linear transformation is: Permeability Values 0.01 0.05 X -min val max val min Val Y = Linear Transformed Values 0.000000 0.oooo20 where Y is the normalised permeability value Xis the actual permeability value minvul is the minimum permeability value in the data set m w u l is the maximum permeability value in the data set 1 5 100 15