Reservoir simulation is essential for various reservoir engineering processes such as history matching and field development plan optimization but is typically an intensive and time-consuming process. The aim of this study is to compare various deep-learning algorithms for constructing a machine-learning (ML) proxy model, which reproduces the behavior of a reservoir simulator and results in significant speedup compared to running the numerical simulator.
Initially, we generate an ensemble of realizations via the reservoir simulator to train the different ML algorithms. The data set consists of a comprehensive set of uncertainty parameters and the corresponding simulation data across all wells. The system utilizes recent advances in deep learning based on deep neural networks, convolutional neural networks, and autoencoders to create machine-learning-based proxy models that predict production and injection profiles as well as the bottomhole pressure of all wells. Thus, the proposed workflows replace the time-consuming simulation process with fast and efficient proxy models.
In this work we provide a comparative study of various ML-based algorithms utilizing deep neural networks and convolutional neural networks for constructing a surrogate reservoir model. The trained models can simulate the behavior of the physics-based reservoir simulator by correlating uncertainty parameters to various history-matched reservoir properties. The algorithms were tested on a mature oilfield with a notable number of wells and several decades of production and injection data. We analyze the performance of each ML approach and provide recommendations on the optimal one.
The best performing workflow for building the ML proxy model consists of two steps. The first step uses stacked autoencoders to learn a low-dimensional latent space representation of the highly dimensional simulation data. This step allows to reduce the complexity of predicting the simulation data and enhances the prediction quality. The following step constructs an ML model to predict the latent space features from input uncertainty parameters and produces highly accurate results.
Reservoir simulation is of paramount importance for various reservoir engineering workflows. Traditional approaches require running physics-based simulators for multiple iterations, which results in time-consuming and labor-intensive processes. We implement and compare several deep-learning-based methods to construct ML proxy models that automate and remarkably reduce the runtime of the reservoir simulation process.