Deep Learning–Based Production Forecasting and Data Assimilation in Unconventional Reservoir

Abstract

Summary Developing unconventional reservoirs such as shale oil is vital for fulfilling the need for energy consumption in the world. Oil production from shale reservoirs is still the most complicated and uncertain phenomenon because of its complex fracture networking, low matrix porosity, and permeability. Production forecasting is crucial for decision-making and tactical exploitation of subsurface resources during production. Traditional methods, such as the Arps decline model and reservoir simulation methods, face significant challenges in forecasting hydrocarbon production due to the highly nonlinear and heterogeneous nature of rocks and fluids. These methods are prone to substantial deviations in forecasting results and show limited applicability to unconventional reservoirs. Therefore, it is essential to improve the production forecasting capability with the help of a data-driven methodology. The data set for modeling is collected from two prominent shale oil-producing regions, the Eagle Ford and the Bakken. The Bakken data set is used to train and test the models, and the Eagle Ford data set is used to validate the model. The random search method was used to optimize the model parameters, and the window sliding technique was used to find a suitable window size to predict future values efficiently. The combination of different deep learning (DL) methods has designed a total of six hybrid models: gated recurrent unit (GRU), long short-term memory (LSTM), and temporal convolutional network (TCN). These models can capture the spatial and temporal patterns in the oil production data. The results concluded that the TCN-GRU model performed best statistically and computationally compared with other individual and hybrid models. The robust model can accelerate decision-making and reduce the overall forecasting cost.