Perpustakaan Digital ITB

In conventional Pressure Transient Analysis (PTA), engineers must visually identify the correct well test model from a log-log plot and then perform iterative regression matching to estimate the reservoir parameters. This process is time-consuming and dependent on the engineer's experience. Previous studies have applied a Convolutional Neural Network (CNN) to automate the model identification step, but the parameter estimation still has to be done manually from scratch. To address this gap, a deep metric learning approach is proposed. A large synthetic dataset consisting of log-log diagnostic plots is generated using analytical Laplace-domain solutions combined with Stehfest numerical inversion, covering four scenario classes: Homogeneous Infinite, Homogeneous One Fault, Dual Porosity Infinite, and Dual Porosity One Fault. The parameters are sampled using a Design of Experiments (DoE) methodology to ensure uniform coverage across the parameter space. A Siamese Neural Network (SNN) is trained to project these plots into a 128-dimensional embedding space where visually similar curves are placed close together. A k-Nearest Neighbors (k-NN) algorithm then retrieves the most similar curves from a reference database to simultaneously predict the scenario class and estimate the parameters through neighbor averaging. The model achieves a classification F1-Score of 0.986 on the test dataset. Parameter estimation accuracy is evaluated using Normalized Absolute Error (NAE), with the model showing generally low errors for permeability, skin factor, and wellbore storage. Higher deviations are observed for boundary distance and storativity ratio, which is consistent with the known non-uniqueness problem in well testing where certain parameter combinations produce visually identical curves. A test case validation confirms the model's applicability, with estimation errors consistently below 2% for permeability, below 4% for skin factor, and below 6% for wellbore storage across all four test cases. The proposed system serves as an initial guessing tool to provide engineers with informed starting values, which can be refined through analytical matching in commercial software.