Automatic well-testing model diagnosis and parameter estimation using artificial neural networks and design of experiments

Abstract

The well-testing analysis is performed in two consecutive steps including identification of underlying reservoir models and estimation of model-related parameters. The non-uniqueness problem always brings about confusion in selecting the correct reservoir model using the conventional interpretation approaches. Many researchers have recommended artificial intelligence techniques to automate the well-testing analysis in recent years. The purpose of this article is to apply an artificial neural network (ANN) methodology to identify the well-testing interpretation model and estimate the model-related variables from the pressure derivative plots. Different types of ANNs including multi-layer perceptrons, probabilistic neural networks and generalized regression neural networks are used in this article. The best structure and parameters of each neural network is found via grid search and cross-validation techniques. The experimental design is also employed to select the most governing variables in designing well tests of different reservoir models. Seven real buildup tests are used to validate the proposed approach. The presented ANN-based approach shows promising results both in recognizing the reservoir models and estimating the model-related parameters. The experimental design employed in this study guarantees the comprehensiveness of the training data sets generated for learning the proposed ANNs using fewer numbers of experiments compared to the previous studies.