Automatic Recognition of Homogeneous and Heterogeneous Reservoirs Using Deep Learning Technique

Abstract

Abstract Reservoir analysis and characterization primarily rely on well test data to indirectly infer crucial reservoir parameters, including wellbore storage, permeability, storativity ratio, and initial reservoir pressure, by examining pressure responses from downhole. While prior studies have showcased machine learning algorithms’ effectiveness in accurately classifying reservoir models with over 90% accuracy, these efforts often have been limited to fewer than ten distinct models. This study introduces a streamlined convolutional neural network (CNN)-based architecture for deep learning, extending its capabilities to classify twenty diverse reservoir models. Pressure transient signals were generated for these models, encompassing both homogeneous and dual porosity reservoirs, each exhibiting ten unique boundary types, including infinite acting, circular sealing, constant circular pressure, single sealing, single constant pressure, angular sealing, channel sealing, channel constant pressure, U-shaped sealing, and rectangular sealing. Utilizing 300 pressure signals for each model, we harnessed well-test analysis software to simulate these signals. To accommodate variations in real field data duration compared to simulated data, we employed an outer product transformation of pressure transient and pressure derivative data, converting it into a grayscale image with values ranging from 0 to 1 and resized to 30×40 dimensions. This transformation preserved crucial patterns, regardless of magnitude and duration. Optimizing the CNN's hyperparameters, including filter number, size, max-pooling, and stride, was achieved through Bayesian optimization, resulting in a five-level deep ResNet architecture. Our model evaluation adopted an 85-15 data split for training and testing. The network's performance was assessed using key metrics, yielding impressive results, with an overall accuracy of 93.3%, sensitivity of 93.4%, specificity of 99.7%, and an F1 score of 93.2%. Furthermore, the model demonstrated consistent accuracy above 90% for fifteen of the twenty reservoir models, even when tested with diverse magnitude and duration pressure signals. This study's innovative approach highlights the effectiveness of a simplified CNN architecture in accurately classifying a broad range of reservoir models, offering reduced training time without compromising accuracy. These findings lay the foundation for future research aimed at further advancing deep learning's potential in reservoir characterization, promising valuable contributions to the field.