Rock Type and Permeability Prediction of a Heterogeneous Carbonate Reservoir Using Artificial Neural Networks Based on Flow Zone Index Approach

摘要

The Cretaceous carbonates of Sarvak formation formed large hydrocarbon reservoirs in the South-west region of Iran. The studied field is a tight carbonate reservoir in which several exploration wells have been drilled, and is in the process of development. Since, only few wells have core data, therefore it was decided to integrate the available core and log data using new methods that describe the carbonates heterogeneity more precise. 3D modeling of permeability is an essential part of building robust dynamic model for proper reservoir management and making reliable predictions. A good definition of reservoir rock types (RRT) could relate somehow better geological modes to dynamic models. Rock typing by flow-zone-index (FZI) and rock-quality-index (RQI) values proved to be an effective technique to develop porosity-permeability transforms for RRTs in a reservoir model. RRTs were defined based on the core derived FZI through some mathematical and statistical approaches. Permeability estimation using artificial neural network approach (ANN) was then made through a two-step process. In the first step, FZI log was estimated from a trained neural network using the standard suite of logs as input (Gamma ray, Sonic, Density, Neutron porosity) and FZI-core as output in a subset of cored wells (Key wells). In the second step, individual trained neural networks implemented porosity-log and FZI-log from the first step to predict permeability-log for each RRT. Validation of the predictive capability of the method in two cored wells (Blind-test wells) that are located in the field proved the estimation technique to be robust and was found to be valuable to supplement core data in the prediction of log- permeability in the entire reservoir wells. For the sake of comparison between the result of this work and the work which was based on the integration of sedimentological, petrographical, and diagenetic study, the results were found to be in good agreement for most of the log interval. However, the predictions of the ANN approach in the regions where core data are not available are better and it follows the log property variation logically.