Deep Convolutional Neural Network (DCNN) Models for Accurate Analysis and Labeling of Complex Reservoir Temporal Data Sequences

Reservoir engineering relies heavily on understanding patterns and signatures in temporal data to inform field development and management decisions. However, classical data analysis and mining methods often struggle to extract meaningful insights from complex, high-dimensional data. This paper proposes a novel approach using Deep Convolutional Neural Networks (DCNNs) to analyze and label temporal reservoir data, enabling more accurate and efficient decision-making. DCNNs are used in many applications such as image classification, facial recognitions, and medical image analysis. In this work, DCNNs are extended to applications in oil and gas to recognize patterns, identify signatures, classify, and label the reservoir temporal data. Temporal reservoir data (e.g., production/injection data, pressures, water-cut and gas-oil ratios) exhibits complex, nonlinear patterns and variable signatures, making it challenging to analyze. DCNN models were developed to work with distinct vector types (e.g., Water and Oil Production Rates and Water-Cut) to predict labels specific to each vector type. For example, in the case of the Oil Production Rates vector, the model predicts labels such as ‘Under producing’, ‘Late matching’, or ‘Matching’. The development of the DCNN models involved utilizing hundreds of images of temporal data to train, validate and test the neural networks. Those images are in the form of line plot images, which were previously labeled. This process also involves data preprocessing, feature extraction, and training using a combination of convolutional and fully connected layers to learn complex patterns in the data. This work utilized 80 realizations of NORNE and BRUGGE benchmark models to demonstrate the efficiency of Deep Convolutional Neural Network (DCNN) models in analyzing reservoir temporal data. Experiments with different numbers of images showed the robustness of the models, with accuracy improving from 89%, using 3900 images, to 97% using 250,000 images. This also demonstrates the models’ ability to learn complex data patterns with high accuracy. The approach enabled the DCNN models to perform independently, without relying on external labeling processes. This has significantly reduced the time, effort, and resources required by engineers, as compared to classical data analysis and mining methods. Additionally, the models are field-independent, allowing them to be applied across different reservoir models and fields. This enables DCNN to evaluate wells and field performance during processes involving reservoir models’ calibration and history matching. This study presents a novel application of Deep Convolutional Neural Networks (DCNNs) in reservoir engineering, leveraging their pattern recognition capabilities to analyze temporal data. By extending DCNNs to oil and gas applications, this research provides a new framework for an accurate temporal data labeling, enabling optimized model calibration and history matching processes, as well as evaluating field development scenarios.