Data Driven Approach for Predicting Natural Gas Outputs from Unconventional Oil Fields Using Machine Learning Techniques

Abstract The increasing global demand for energy has led to a rise in the exploration and production of unconventional oil fields. As conventional resources become more challenging to access, unconventional oil fields such as poorly consolidated or unconsolidated formations are becoming ideal sources of natural gas production. However, predicting natural gas production from these complex, high viscosity oil fields remain a challenging task due to the heterogeneity of the geological formations and the varying production rates. This research focuses on the integration of advanced machine learning (ML) techniques to address these challenges, providing a data-driven approach for improving natural gas forecasting in unconventional oil fields. By leveraging historical production data, reservoir characteristics, and operational parameters, machine learning models, such as Linear regression, decision trees, random forests (RF), gradient boosting methods and artificial neural network (ANN) are employed to identify patterns and relationships within the data that influence gas production. Emphasis is placed on optimizing feature selection, data preprocessing, and model tuning of over 15,000 data points to improve the prediction accuracy and then validated using various performance metrics, including coefficient of determination (R2) and root mean squared error (RMSE). The Categorical Boosting (CatBoost) regression model achieved a R2 score of 0.956 with a MSE of 0.1205 while the Extreme Gradient Boosting model (XGBoost) and Random Forest model coming close with a R2 score of 0.953 and 0.949 respectively. This data-driven approach not only enhances short and long-term production forecasting but also offers insights into optimized recovery strategies and reservoir management. This study demonstrates how machine learning algorithms fit into the natural gas production prediction flow, evaluates their efficiency, and the need for cross-disciplinary approach in integrating machine learning with existing reservoir engineering practices in the industry. In conclusion, this work demonstrates the potential of leveraging Big Data and ML techniques in real-time to reduce operational risks, improve decision making and revolutionize natural gas production forecasting, contributing to more efficient and sustainable energy exploration.