Automatic recognition of oil reservoir models from well testing data by using multi-layer perceptron networks

Hydrocarbon reservoirs are complex heterogeneous media, and their parameters are usually identified indirectly through using well testing techniques. The well testing is basically conducted through creating a flow disturbance in the well and recording the related response of the bottom-hole pressure. This technique provides the needed data for quantitative analysis of the reservoir parameters and reservoir characterization. The well testing method consists of two stages: (1) the recognition of the reservoir model, and (2) the parameter estimation. The aim of this study is to apply the artificial neural network (ANN) to the recognition of the reservoir model. The structure of the neural network used in this work is a multi-layer perceptron (MLP) network. The required training and test data sets have been generated by using the analytical solutions of commonly-used reservoir models. Eight important reservoir models considered in this study include homogenous and dual porosity reservoir models with different outer boundaries such as no flow, constant pressure, infinite acting and single sealing fault boundaries. The mean relative errors (MRE) and the mean square errors (MSE) of the test data have been used for determining the number of neurons in the hidden layer. The required CPU time for the training of the proposed network has also been utilized for selection of the most suitable training algorithm. A two-layer MLP network with twelve neurons in its hidden layer has been designed as the best configuration. The scaled conjugate gradient method has been chosen as the training algorithm. The performance of the proposed ANN has been examined by the actual field data in addition to simulation noisy and noiseless data sets. The results indicate that the proposed two-layer MLP network can identify the reservoir models with an acceptable accuracy.