Machine learning techniques have garnered significant attention in various engineering disciplines due to their potential and benefits. Specifically, in reservoir numerical simulations, the core process revolves around solving the partial differential equations delineating oil, gas, and water flow dynamics in porous media. Discretizing these partial differential equations via numerical methods is one cornerstone of this simulation process. The synergy between traditional numerical methods and machine learning can enhance the precision of partial differential equation discretization. Moreover, machine learning algorithms can be employed to solve partial differential equations directly, yielding rapid convergence, heightened computational efficiency, and accuracies surpassing 95%. This manuscript offers an overview of the predominant numerical methods in reservoir simulations, focusing on integrating machine learning methodologies. The innovations in fusing deep learning techniques to solve reservoir partial differential equations are illuminated, coupled with a concise discussion of their inherent advantages and constraints. As machine learning continues to evolve, its conjunction with numerical methods is poised to be pivotal in addressing complex reservoir engineering challenges.
Summary: During sucker rod pump production, there is a commonly seen problem of class imbalance, which refers to the differences in the amount of data accumulated under different working conditions. This problem in rod pump diagnosis can lead to unsatisfactory classification results of surface dynamometer cards under working conditions with fewer samples. Therefore, this study adopts the conditional generative adversarial nets (CGANs) improved by mini-batch method to address the problem of class imbalance. CGAN is an efficient method of multiclass data generation, which learns the properties of dynamometer cards by training the generator and discriminator networks. CGAN is modified by mini-batch strategy to avoid mode collapse and enable the interaction among input samples of discriminator, so that the generated dynamometer cards can be much more diverse. Results show that the shapes of generated dynamometer cards are basically consistent with those of real samples, and the structural similarity (SSIM) among the generated samples decreases, indicating that the generated dynamometer cards have more types of shape. Meanwhile, as the generated dynamometer cards become more diverse, their differences with real samples in data distribution are reduced, according to the calculation of sliced Wasserstein (SW) distance. Based on real and generated dynamometer cards, we developed the classifiers for working condition diagnosis of rod pump through convolutional neural network (CNN). The classification results of the validation set indicate that without the mini-batch method, the recall of generated categories for pump hitting down and leakage has increased by 12 and 5.3%, respectively; in contrast, with the mini-batch method, the recall has increased more obviously by 7, 24, and 2%, respectively, for gas lock, pump hitting down, and leakage. Our research results have demonstrated that the proposed method can effectively solve the problem of insufficient data accumulation in the oil field.
The reservoir characterization aims to provide the analysis and quantification of the injection-production relationship, which is the fundamental work for production management. The connectivity between injectors and producers is dominated by geological properties, especially permeability. However, the permeability parameters are very heterogenous in oil reservoirs, and expensive to collect by well logging. The commercial simulators enable to get accurate simulation but require sufficient geological properties and consume excessive computation resources. In contrast, the data-driven models (physical models and machine learning models) are developed on the observed dynamic data, such as the rate and pressure data of the injectors and producers, constructing the connectivity relationship and forecasting the productivity by a series of nonlinear mappings or the control of specific physical principles. While, due to the “black box” feature of machine learning approaches, and the constraints and assumptions of physical models, the data-driven methods often face the challenges of poor interpretability and generalizability and the limited application scopes. To solve these issues, integrating the physical principle of the waterflooding process (material balance equation) with an artificial neural network (ANN), a knowledge interaction neural network (KINN) is proposed. KINN consists of three transparent modules with explicit physical significance, and different modules are joined together via the material balance equation and work cooperatively to approximate the waterflooding process. In addition, a gate function is proposed to distinguish the dominant flowing channels from weak connecting ones by their sparsity, and thus the inter-well connectivity can be indicated directly by the model parameters. Combining the strong nonlinear mapping ability with the guidance of physical knowledge, the interpretability of KINN is fully enhanced, and the prediction accuracy on the well productivity is improved. The effectiveness of KINN is proved by comparing its performance with the canonical ANN, on the inter-well connectivity analysis and productivity forecast tasks of three synthetic reservoir experiments. Meanwhile, the robustness of KINN is revealed by the sensitivity analysis on measurement noises and wells shut-in cases.
1 September 2022 in Journal of Energy Resources Technology