A Combined Neural Network Forecasting Approach for CO2-Enhanced Shale Gas Recovery

Summary Intensive growth of geological carbon sequestration has motivated the energy sector to diversify its storage portfolios, given the background of climate change mitigation. As an abundant unconventional reserve, shale gas reservoirs play a critical role in providing sufficient energy supply and geological carbon storage potentials. However, the low recovery factors of the primary recovery stage are a major concern during reservoir operations. Although injecting CO2 can resolve the dual challenges of improving the recovery factors and storing CO2 permanently, forecasting the reservoir performance heavily relies on reservoir simulation, which is a time-consuming process. In recent years, pioneered studies demonstrated that using machine learning (ML) algorithms can make predictions in an accurate and timely manner but fails to capture the time-series and spatial features of operational realities. In this work, we carried out a novel combinational framework including the artificial neural network (ANN, i.e., multilayer perceptron or MLP) and long short-term memory (LSTM) or bi-directional LSTM (Bi-LSTM) algorithms, tackling the challenges mentioned before. In addition, the deployment of ML algorithms in the petroleum industry is insufficient because of the field data shortage. Here, we also demonstrated an approach for synthesizing field-specific data sets using a numerical method. The findings of this work can be articulated from three perspectives. First, the cumulative gas recovery factor can be improved by 6% according to the base reservoir model with input features of the Barnett shale, whereas the CO2 retention factor sharply declined to 40% after the CO2 breakthrough. Second, using combined ANN and LSTM (ANN-LSTM)/Bi-LSTM is a feasible alternative to reservoir simulation that can be around 120 times faster than the numerical approach. By comparing an evaluation matrix of algorithms, we observed that trade-offs exist between computational time and accuracy in selecting different algorithms. This work provides fundamental support to the shale gas industry in developing comparable ML-based tools to replace traditional numerical simulation in a timely manner.