Abstract
Abstract
The understanding of fracture distributions plays a critical role in managing fractured reservoirs as they govern early water/CO2 breakthrough, impact sweep efficiency, and determine production behaviors. However, traditional simulation-based approaches, such as history matching, encounter significant difficulties in accurately predicting fracture distributions, and high-fidelity simulations can be computationally prohibitive. This paper proposes a comprehensive machine learning-based workflow to effectively characterize and describe fracture distributions for unconventional reservoir models.
The proposed workflow has four components. Firstly, a single fracture parameterization is implemented, utilizing four fracture parameters: fracture initiation point, length, angle, and azimuth. Secondly, a Variational Autoencoder (VAE) is employed for fracture map parameterization. The encoder maps a high-dimensional fracture distribution map to a low-dimensional latent space, and the decoder reconstructs the fracture distribution map from the reduced latent dimension to the full reservoir dimension. Thirdly, a neural network is utilized for fracture distribution prediction, establishing a regression relationship between latent variables and production data. Finally, a nearest-neighbors selection is achieved by applying principal component analysis (PCA) in 2D principal coordinates for quantifying uncertainty.
The efficacy of the proposed workflow is demonstrated in a 2D synthetic case and subsequently applied to the 3D benchmark case. A total of 5,000 fractured permeability realizations are generated by randomly selecting the four fracture parameters. The values for these parameters are generated based on a normal distribution. Each realization has a unique fracture distribution. These realizations are split into training (4,500), validation (250) and testing (250) sets. The VAE model is trained on the training set first. Then the best model was selected using the validation set, and finally tested on the testing set. The trained VAE decoder serves as a fracture generator. A total of 200 latent variables are selected to represent the latent fracture distribution and fed to the decoder to reconstruct the fracture maps. To predict an unknown fracture distribution given only observed production data, we establish regression models between the production data and latent variables. The regression models are neural network models trained on the production data and the latent vectors of the training set. In the prediction stage, the observed production data was fed to the regression models to predict the latent vectors. Then the latent vectors were passed to the trained VAE decoder to predict the latent fracture maps. Finally, to account for the geological uncertainty, we applied the nearest neighbor selection to select multiple realizations from the training and validation set as the results.
The comprehensive data-driven workflow presented in this paper not only offers an efficient and effective way for fracture parameterization and prediction, but also demonstrates the practical feasibility in field case study.