Horizontal Well Performance Predictions for Carbon Storage Using Deep Neural Networks

Abstract This work aims to develop a deep learning-based proxy model for near real-time prediction of horizontal well performance using available vertical well data for field-scale carbon capture and storage (CCS) projects. Traditional full-physics reservoir simulations for Area of Review (AoR) and well injectivity assessments are computationally expensive for field-scale high resolution models, particularly for scenario-based injectivity assessments considering geologic uncertainty. We propose a deep neural network architecture combining Variational Autoencoder (VAE) with feed-forward regression model to rapidly predict horizontal well injectivity and AoR, enabling faster assessment and decision-making for field development strategy in CCS applications. An important aspect in large-scale applications of deep learning models is rapid generation of comprehensive training data. We utilize a Pseudo-Steady-State (PSS) based multi-resolution model (PSS-SIM) to accelerate compositional simulations and training data generation. The PSS-SIM coarsens the reservoir model along PSS-pressure contours, enabling rapid simulations for large-scale field applications. As a first step of the proposed workflow, a detailed sensitivity analysis is conducted to identify the key uncertainty parameters, which are influential to the AoR and injectivity. Next, Latin Hypercube Sampling (LHS) is utilized to generate training data set accounting for the uncertainty range of the identified influential parameters. The training data consists of injectivity and AoR for horizontal as response variable and predictor variables include injectivity and AoR of vertical wells, and the horizontal well design (aspect ratio and lateral length). The VAE compresses high-dimensional AoR images into lower dimensional latent variables, while a feed-forward neural network-based regression model predicts horizontal well injectivity and latent variables from the vertical well performance and the horizontal well design. We first validate the accuracy of the PSS-SIM by comparing against fine-scale simulations for both vertical and horizontal injection well scenarios. The calibrated PSS-SIM achieves a 32-fold acceleration in the simulation CPU time while maintaining sufficient accuracy. Next, sensitivity analysis is conducted to evaluate the impact of horizontal well parameters on the AoR and injectivity index, including horizontal well lateral length, kv/kh in different zones, and zonal location of the horizontal well. The sensitivity study identifies the key influential parameters: kv/kh in the injection zone, lateral length, and well placement location. Using PSS-SIM, we generate 600 training samples to properly capture the variations in these parameters. The proposed deep neural network is trained with a loss function integrating AoR image reconstruction, latent variable estimation, and injectivity index prediction. The neural network model demonstrates sufficient accuracy in AoR image and injectivity index prediction. The proposed approach enables rapid training data generation, AoR and injectivity predictions in near real-time for horizontal wells in CCS applications based on available data from vertical well scenarios. The core innovation is the deep neural network for image compression combined with feed-forward regression to predict AoR images and injectivity indices in near real-time for horizontal wells. We employ PSS-SIM that accelerates compositional simulations by orders of magnitude, significantly speeding up training data generation. The integrated framework enables fast and accurate well performance predictions and facilitates seamless integration into the platform for CCS applications.