Predicting Future Heat Outputs from Enhanced Geothermal System Utilizing Machine Learning Approach

Abstract

Abstract The Earth is a vast energy reservoir. The U.S. Department of Energy estimates that harnessing just 0.1% of the Earth's geothermal energy can power humanity for 2 million years. The energy sector has shown a significant interest in geothermal energy owing to its advancements in renewable energy, environmental friendliness, and widespread accessibility. An improved geothermal system (EGS) efficiently extracts heat from deep hot dry rock (HDR). However, EGS is battling to ensure safe drilling and appropriate fracturing to extract heat potential. Essential aspects to evaluate are deformation and fracture face damage during induced fracturing in order to extract heat energy from HDR, due to its heterogeneities. This study examines and predicts future heat outputs from EGS utilizing machine learning. The UTAH FORGE well, 16B (78)-32, provided the well logs and petrophysical characteristics. The single-well data was divided into three categories: training, testing, and validation, with a 70:20:10 ratio. The model was built using eleven well-log variables in total, including anisotropy in heat, density, porosity, Poisson ratio, compressional and shear travel times, and SP and GR. Machine Learning model (ML), Decision Tree (DT) and Random Forest (RF) model were constructed, and an optimization technique was employed to ascertain the hyperparameters of the ideal model for heat production prediction. The pair plot indicates that there is no discernible noise present in the recorded data, and the correlation matrix illustrates a perfect correlation (unity) between temperature and depth. The machine learning model exhibited outstanding performance in forecasting the future temperature of the geothermal reservoir. Both Random Forest (RF) and Decision Tree (DT) models displayed exceptional accuracy, achieving R2 scores exceeding 98% with RMSE values below 3%. Particularly, the Random Forest model surpassed traditional approaches, achieving an accuracy of approximately 99.7%. These results suggest that these models remain capable of generating reliable and useful projections.