Comparison of machine learning algorithms for the development of a forecasting tool for cyclic gas injection in hydraulically-Fractured wells

Single-well cyclic gas injection (cyclic pressure pulsing) is an enhanced oil recovery (EOR) method that includes injecting different types of gases that soaks into formation. In hydraulically-fractured wells, surface area of fracture can help the gas-diffusion process to be more effective. After the injected gas penetrates into the rock matrix, increased pressure and interaction of injected gas with the oil can result in improved production rates. As in all other EOR applications, performance-forecasting is critical to properly design the operational aspects of this application. In situations where a reliable numerical model or reservoir simulator is not accessible, machine learning methods can offer an alternative approach for forecasting the performance in a practical manner. When the dataset used for training is generated from numerical simulation runs, machine learning model serves as a proxy to the reservoir simulator by accurately predicting the performance variables from a set of input parameters. This eliminates the necessity to have access to a reservoir simulator and helps to conduct comprehensive analyses in a timely manner. In this study, various machine learning algorithms were tested to develop predictive models for the cyclic injection process by using a dataset obtained from numerical simulation runs. To design machine learning models, we used the provided database with 5 operational parameters: injection rate, injection time, soaking time, cycle rate, the composition of three injected gases (CO2, N2, CH4) and output as the economic efficiency for each year derived by simulations. Based on the given simulations data, three different machine learning models were constructed: Random Forest, Neural Network, and Multiple Linear Regression. Firstly, data is analyzed by the visualization of input variables and importance identification to observe the correlation between each other and find similar patterns. Secondly, models were created based on input and output variables. Finally, improvement of prediction performance, accuracy optimization, and error computational techniques are applied to each of the three methods using R tool. Among the three algorithms, validation performance of random forests was the best with a coefficient of determination of 0.93.