Enhancing Pressure Transient Analysis with Automatic Model Identification: A Machine Learning Approach

Abstract

Abstract Over its historical trajectory, Pressure Transient Analysis (PTA) has experienced a notable evolution, commencing with early techniques like straight-line analysis and type curve matching. The mid-20th century witnessed the introduction of pressure derivatives, enhancing the sophistication of reservoir behavior interpretation. Deconvolution methods took center stage in the late 20th century, offering intricate insights into wellbore and reservoir responses. In the contemporary era of PTA, log-log analysis has become the norm, featuring the plotting of pressure and its derivatives on logarithmic scales. Recent strides in the field concentrate on the integration of automation and machine learning to expedite PTA processes. Our methodology enhances the Pressure Transient Analysis (PTA) process by leveraging a framework based on triplet loss. This architecture seamlessly integrates Convolutional Neural Network (CNN) layers, providing robust feature extraction capabilities for the automated analysis of pressure transient data. The model is trained using simulated experimental data generated through a systematic Design of Experiments (DOE) approach. This involves incorporating the ten most prevalent interpretation scenarios, encompassing well, reservoir, and boundary model types. For each model type, critical parameters such as permeability, horizontal well length, skin factor, and distance to the boundary are systematically sampled, resulting in the computation of 800 distinct pressure derivative responses. The triplet loss framework adopts a self-supervised training strategy, where anchor, positive and negative pairs are dynamically generated from the simulated training dataset. The loss function encourages the network to reduce the distance between the anchor and positive examples while increasing the distance between the anchor and negative examples. The experimental analysis revealed that the actual model class consistently ranked high among the top classes. The model exhibits an accuracy rate of 90% in providing the top-ranked model recommendations when evaluated on 100 samples derived from the 8 interpretation scenarios. Having prior knowledge about the most probable well test models at the top ranks diminishes the manual effort required for analysis. This approach can expedite the identification of the pressure derivative response associated with specific combinations of well, reservoir, and boundary models, leading to the generation of models with reduced reliance on user interaction. The methodology streamlines the recognition of models for interpretation engineers, enabling faster integration with additional information from diverse sources like geophysics, geology, petrophysics, drilling, and production logging.