Coal gas content prediction on Kinik coalfield, Soma Basin with machine learning methods

Abstract

Coal has been used by humanity since ancient times, becoming widespread with steam engines, and is now a complex energy source that is beginning to be replaced by alternatives. While coal is used as a direct source of energy, it also serves as a source rock that produces various fluids, primarily carbon-based gases. Its multiple functions and the presence of various quality and quantity reserves in many countries fundamentally extend the lifetime of coal. Moreover, the presence of methane and other natural gas components in coal makes it capable of contributing to natural gas reserves as an effective alternative when coal is phased out. This study introduces a novel data-driven methodology for interpreting the nonlinear challenge of analyzing the total desorbed gas content in coal seams. The investigation is centered on a low-rank coal reserve situated in the Kınık coalfield, where the United States Bureau of Mines (USBM) direct desorption method was employed to project the total desorbed gas content for underground mining operations. Utilizing core samples obtained during the reserve and gas content analysis, machine learning models were developed. These models were trained with coal properties data, including depth, moisture, ash, volatile matter, and calorific value, in correlation with the total desorbed gas content. Various machine learning algorithms, namely multiple linear regression, support vector machine, and artificial neural network, were utilized to predict the total desorbed gas content in the Kınık coalfield. Hyperparameter tuning was applied to optimize the machine learning models, and the most effective model was chosen based on its regression accuracy and computational efficiency. The raw data analysis, facilitated by pairplot, revealed associations between parameters and their direct influence on the total gas content in coal seams. Sensitivity analysis was performed to assess the impact of coal properties on total desorbed gas content. The selected model was then applied to predict the total desorbed gas content at a specific location in the coalfield. The study's outcomes offer valuable insights and recommendations for the analysis of unconventional reservoirs and the prediction of petrophysical systems using machine learning techniques. In essence, this research underscores the potential of machine learning in tackling nonlinear challenges within the geological domain and proposes a promising avenue for future investigations in this field.