Probabilistic Geothermal resources assessment using machine learning: Bayesian correction framework based on Gaussian process regression

Abstract

Accurate determination of geothermal resources in the study area is of great importance for effective exploration and development. In the present study, geothermal resources in Qianguo County, Jilin Province, China, were selected as the research object, and we used the random forest sensitivity analysis method to determine the five most influential parameters among the nine parameters involved in the volumetric method: geothermal reservoir thickness (H), geothermal reservoir temperature (T), geothermal reservoir area (A), thickness porosity (Φ), and rock density (ρw). Among them, the uncertain parameters H and T show a significant positive effect on the output results. These parameters and geothermal resources estimated by volumetric method based on Monte Carlo were used as input parameters in the Gaussian process simulation, and the squared exponent (SE) with fewer hyperparameters was selected as the Gaussian process regression kernel function. During the analysis, six groups of different hyperparameter combinations were set, and the Earth Mover's distances (EMD) was applied to compare the priori distribution and simulation results with the posterior distribution (true distribution). The simulation results that are closest to the target distribution of geothermal resources were used as the final results of geothermal resources calculation. Based on the obtained results, it was concluded that the Qingshankou and Quantou formations located in the southern Changling sag have good geothermal development potential which the geothermal resources are 5.80 × 1010 kJ·km−2 and 6.22 × 1010 kJ·km−2, and the probability of occurrence is 4.65%, and 3.69%.