Data-Driven Reservoir Modeling using Recurrent Neural Network and Physics-Based Network Model

Abstract

We present efficient data-driven reservoir model workflows for a mature oil field involving large-scale CO2 Water Alternating Gas (WAG) injection. The CO2 WAG injection is conducted in more than two hundred wells in the entire field, and the operation area is spread throughout the field. Therefore, it is computationally prohibitive to implement history matching or optimization using full-field reservoir models. The objective of this study is to develop efficient data-driven approaches to optimize the CO2 WAG operation and maximize oil recovery from the reservoir. The proposed workflows are useful for predicting future production rates and understanding the reservoir connectivity between producers and injectors. We propose two different types of approaches. First, deep learning algorithms are utilized to develop efficient data-driven reservoir models. Long Short-Term Memory (LSTM) is a special kind of neural network architecture and has been successfully applied to many sequential and time series problems. We formulate time series problems of the production and injection histories, and the LSTM algorithm is used to forecast the future production rate and to estimate the reservoir connectivity. Second, we utilize a physics-based data-driven reservoir model, the 1D network model. The 1D network model characterizes a reservoir by a network grid system, which connects each producer injector pair via a series of 1D grid cells. Numerical reservoir simulators compute the solution of the network grid system. History matching is implemented by Ensemble Smoother with Multiple Data Assimilation (ESMDA), and a streamline-based rate allocation optimization is implemented based on the calibrated network model. The LSTM reservoir modeling workflow was validated using synthetic reservoir cases. It showed reasonable performance on production rates forecasting and reservoir connectivity estimation. Then, we successfully implemented this approach for a real field application. The 1D network model provided suitable history matching results for the entire field application of the mature oil reservoir. Moreover, a streamline-based rate allocation optimization was implemented, and it provided improved oil recovery from the reservoir.