Summary
This paper describes use of backpropagation neural network (BPNN) technique to predict reservoir permeability using conventional well log data. The technique is demonstrated with an application to the Ravva oil and gas field, offshore India. The Ravva field reservoirs are middle Miocene age nearshore marine sandstones that are often laminated to thinly interbedded with shale. The use of conventional permeability-porosity crossplots to predict permeability in this field was not successful. The BPNN permeability prediction model ("RAVVANET") was developed from a data set consisting of core permeability and well log data from two early development wells. The model was blind tested using data from a third well which was withheld from the modelling process. The results of this study show that BPNN model permeability predictions are consistent with core analysis results.
Introduction
Permeability, a critical parameter for any field reservoir management is often determined from well log derived variables such as porosity. Often however, porosity and permeability may be independent reservoir properties. If porosity is disconnected, permeability will be low, whereas permeability may be high if porosity is interconnected and effective. Microporosity is generally non-effective and does not significantly contribute to reservoir permeability.
Despite these observations, theoretical relations between permeability and porosity have long been sought. For example, the Kozeny-Carmen theory relates permeability to porosity and specific area of a porous rock with pores treated as an idealised bundle of capillary tubes. This theory, however, treats the highly complex porous medium in a very simple manner and ignores the influence of convergent flow in the pore constrictions and expansions of flow channels. Empirical relations based on the Kozeny-Carmen theory have also been developed which relate permeability to other well logs and/or log-derived parameters such as resistivity, and irreducible water saturation. They may apply either to only the region above the transition zone, or only to the transition zone itself. Due to the limitations of these empirical models, statistical methods have been proposed as a more versatile solution to the problem of permeability estimation.
Statistical regression is widely used to search for relationships between petrophysical properties and well log data. This parametric method which requires the assumption and satisfaction of multi-normal behaviour and linearity must be applied with caution. Size discussed the use and abuse of statistical methods in the geosciences.
A relatively new, non-linear and non-parametric tool called artificial neural networks, or simply neural nets, is becoming increasingly popular in well log analysis. This technique has been applied to permeability predictions. The most commonly used neural net models are backpropagation neural nets (BPNN). Recent comparison studies have shown that BPNN models may be more accurate than conventional methods and statistical regression.
This paper presents an example use of BPNN to determine permeability from well log response in the Ravva field, offshore India. Unlike other studies, this paper describes different stages of model development: data acquisition, selection of inputs, removal of anomalous data, model validation and scope for further development.
In the example, we used core permeability and well log data from two early development wells to develop the RAVVANET, a BPNN permeability prediction model. A third cored well was used as a "blind test" of the model. Performance was evaluated by comparing the predictions with the core data. Following is a brief review of BPNN, followed by a description of Ravva field reservoirs. We will then present the stages of model development.