Artificial neural network tool development for flue gas sequestration in depleted shale oil reservoirs

Abstract

The energy needs of societies are increasing day by day. Oil and natural gas use have become more common with developing technology to meet this demand. However, this consumption also brings some concerns. The negative effects of fossil fuels on the environment, especially greenhouse gas (CO2, CO, N2O, H2S) emissions, have become a cause for worry. Today, the concentration of carbon dioxide in the atmosphere is increasing. Sequestration of flue gas released from factories and power plants is one of the methods used to reduce the rate of greenhouse gases in the atmosphere. In this study, an artificial neural network (ANN) tool that predicts the sequestration of flue gas has been developed. With the developed tool, production estimation can be made for hydraulically fractured shale oil reservoirs with horizontal wells in addition to the flue gas injection time and pressure. The reservoir model used in this study has some key features, such as the stimulated reservoir volume approach, the double porosity double permeability phenomenon and Langmuir adsorption isotherm formulation. After creating the desired reservoir model, minimum and maximum value ranges were taken from the literature for reservoir characteristics and operational parameters. According to these intervals, 10,000 different models with normal distribution were generated randomly and simulated with the CMG GEM simulator. Randomly generated scenario variables and obtained results from the simulation were used in artificial neural network training as input and output values. An artificial neural network, a machine learning method, consists of input, hidden, and output layers. Each layer contains artificial neurons, interconnected like neurons in the biological nervous system, forming the artificial neural network structure. A weight is assigned to the data entering the input layer and processed using an activation function, and output is acquired as a result of the calculations. This study used the hyperbolic tangent function as the activation function. In addition, functional links are used to improve the relationship between input and output. Python and MATLAB programming environments were used to develop the artificial neural network tool. As the training algorithm, the RMSProp function in the Keras library was used in the Python model, and the trainscg and learngdm functions were used in the MATLAB model. In order to minimize the differences arising from the different algorithms used by the models, the ANN structure and the data set used are the same. Accordingly, three hidden layers and 70, 90 and 40 neurons were used in these layers, respectively, in both models. In addition, the data set containing 39 input and 73 output parameters were divided into 80% training, 10% validation and 10% test set. The tools obtained at the end of the training were tested using test sets, and the error rates in the estimations of variables such as shale oil production curves, flue gas injection pressure and injection stopping time were calculated. These error rates were found to be 5.47% for the Python model and 2.54% for the MATLAB model.