4:15 PM - 4:30 PM
[MGI33-08] Development of a deep learning approach to estimate numerical simulation parameters of a hydrothermal system: Evaluation using a two-dimensional model
Keywords:Deep learning, Neural network, MDS, Hydrothermal system, Permeability, Geothermal
For the development of a hydrothermal energy resource, it is important to understand its subsurface temperature distribution. In this study, we proposed a new method for estimating subsurface temperature distributions using deep learning, and investigated the accuracy and characteristics of this proposed method using two-dimensional simulation data. The proposed method first learns the relationship between the simulation parameters (e.g., permeability) and simulated temperatures at well locations (training step), and then estimates the parameters from observed temperature logs (estimation step). The data used in the training step is called training data. In general, the distribution of training data needs to be consistent with the observations; however, a procedure to ensure that this requirement is met has not been established. Therefore, we developed a procedure for visualizing and selecting training data based on multidimensional scaling (MDS), and examined whether the procedure can reduce estimation errors.
In this study, we used multiple sets of two-dimensional training and test data generated using the AUTOUGH2 hydrothermal simulator. For generating the simulation data, formation permeabilities and bottom boundary conditions provided to the simulator were varied. Subsequently, we trained a deep neural network to learn the relationship between the variable parameters and temperatures at pseudo-well locations. In addition, we used MDS analysis to check whether (pseudo) observed temperature logs were included in the range of temperature profiles covered by the training data, and we examined whether parameter estimation errors can be reduced by selecting training data that are most consistent with the pseudo temperature logs.
We found that estimation errors were smallest when using training data generated based on small parameter variabilities. This demonstrates the effectiveness of narrowing down the parameter ranges used to generate the training data in order to improve estimation accuracy. Then, based the MDS analysis results, we compared the estimation reliability between cases where the pseudo temperature logs were consistent and inconsistent with the training data. The error rate for the consistent case was 1.4%, while the error rate for the inconsistent case was 29.5%. Thus, we concluded that the data to be estimated must be included in the range covered by the training data, and the consistency can be assessed based on MDS analysis. Further, the error rate was reduced when selecting the training data based on the MDS analysis. These results demonstrate the effectiveness of using MDS analysis to improve the estimation accuracy of deep learning.
In this study, we used multiple sets of two-dimensional training and test data generated using the AUTOUGH2 hydrothermal simulator. For generating the simulation data, formation permeabilities and bottom boundary conditions provided to the simulator were varied. Subsequently, we trained a deep neural network to learn the relationship between the variable parameters and temperatures at pseudo-well locations. In addition, we used MDS analysis to check whether (pseudo) observed temperature logs were included in the range of temperature profiles covered by the training data, and we examined whether parameter estimation errors can be reduced by selecting training data that are most consistent with the pseudo temperature logs.
We found that estimation errors were smallest when using training data generated based on small parameter variabilities. This demonstrates the effectiveness of narrowing down the parameter ranges used to generate the training data in order to improve estimation accuracy. Then, based the MDS analysis results, we compared the estimation reliability between cases where the pseudo temperature logs were consistent and inconsistent with the training data. The error rate for the consistent case was 1.4%, while the error rate for the inconsistent case was 29.5%. Thus, we concluded that the data to be estimated must be included in the range covered by the training data, and the consistency can be assessed based on MDS analysis. Further, the error rate was reduced when selecting the training data based on the MDS analysis. These results demonstrate the effectiveness of using MDS analysis to improve the estimation accuracy of deep learning.