A Data-Driven Approach for Hydraulic Fracturing Simulation in Shale Based on Time-Series Images of Fracture Propagation

Abstract

Summary Optimizing the hydraulic fracturing design requires efficient and accurate simulation of fracture propagation. However, traditional numerical methods are computationally expensive due to the solution of coupled differential equations, and the simulation accuracy may be reduced by the uncertainty of the input parameters (e.g., rock mechanical and fluid properties). To address these issues, this paper proposes a deep-learning-based approach to improve both the efficiency and accuracy of fracturing simulation in shale. This study develops a novel method integrating Feedforward neural network (FNN) and ConvLSTM to predict fracture propagation in shale. We concatenate the well position images, natural fracture images, and the time-series images of fracture propagation, and use FNN to extract the features of reservoir properties and pumping schedules at each timestep as the input of model. Then, the ConvLSTM network is utilized to extract and fuse features from natural fractures, wellbore locations, and the features extracted by FNN. Data preprocessing techniques are employed improve data quality through cleaning and normalization. Fracture propagation images, wellbore images, natural fracture images, and pumping schedules for hydraulic fracturing were generated using fine-grid hydraulic fracturing simulation. Based on the various settings of different geologic and operational parameters, a dataset with over 1 million samples was established by collecting the fracture propagation image at each frame. The proposed model predicts the fracture morphology images in the next 5 frames based on the fracture propagation history image of the previous 1 frame. The model was evaluated using Structural Similarity Index (SSIM), Mean Squared Error (MSE) and Frame Mean Absolute Error (FMAE). To expedite model training convergence, the Scheduled Sampling technique was incorporated. After 500 iterations of training, the model demonstrated an MSE less than 15×10-5, a maximum SSIM of 0.90, and an average FMAE below 50. In comparison with traditional fracturing simulation using the finite element method, the proposed data-driven method demonstrated a 60% improvement in simulation efficiency. The main value of this work lies in the development of a new data-driven and mesh-free method for predicting fracture morphology, which eliminates the numerical computation issues so that fast and accurate predictions of fracture propagation can be achieved. Without the heavy computational cost in the traditional fracturing simulation, the developed workflow can be integrated with reservoir simulation and optimization algorithms to perform fast and reliable optimization of fracturing design.