9:45 AM - 10:00 AM
[MGI35-04] Estimation of geothermal structure over Japan Island by machine learning incorporating various crustal information with application to geothermal resource assessment
Keywords:Deep Neural Network, Neural Kriging, Stacking Method, critical point
The importance of promoting geothermal power generation has been growing in Japan in order to reduce greenhouse gas emissions, and supercritical geothermal power generation, which generates large amounts of electricity, has attracted particular attention. However, long lead times, development costs, and large risks hinder this promotion, and suitable locations for supercritical geothermal power generation are difficult to be specified. For these situations, it is essential to clarify the geothermal structure from the surface to deep depths over Japan Island. Although the most reliable data for the estimation of the geothermal structure is temperature logging data, both the depth range and amount of the logging data are limited. Therefore, 3D temperature estimation is impossible through usual GIS-based methods such as kriging and spline, which cannot take various crustal information into account and has a strong smoothing effect. To overcome this problem, this study applies Deep Neural Network (DNN) and Neural Kriging (NK), which can flexibly consider such information and reduce the smoothing effect.
At first, we examined what kind of crustal information contributes to increasing the estimation accuracy of geothermal structure by evaluating the feature importance of DNN. The importance of crustal information such as the distribution of active volcanoes, the Curie point depth, rock type, geochemical data of hot springs, temperature distribution at different elevations estimated by the Stacking Method, and geothermal gradient were evaluated as essential factors of the estimation accuracy, in particular, the horizontal temperature distribution, geothermal gradient, and geochemical data of hot springs were the predominant factors.
Next, applicability of NK to all area of Japan Island was examined. The spatial correlation structure of the temperature logging data must be different in each region and between geothermal and non-geothermal regions. Therefore, we constructed a NK model that incorporates the characteristics of the spatial correlation structure in each region. This NK model was proved to increase the extrapolation accuracy and produce smaller error than DNN.
Finally, a three-dimensional geothermal structure analysis was implemented using NK that incorporated all the types of crustal information. A noteworthy result was that the model was able to estimate the geothermal structure from the surface to deep depths over Japan Island from the temperature logging data with limited amount of data points and depth range. In addition, the iso-surface map of the water critical point at 374 °C and 22.1 MPa was constructed by this model, which highlighted promising supercritical geothermal resource areas around Toyoha, Hachimantai, Mt. Kirishima, and Mt. Kurikoma. Furthermore, calculation result of geothermal resource potential revealed that the areas of Toyoha and Mt. Osore in northern Japan are the highest potential in addition to the areas where the geothermal power plants are actually working. Their expected power generation capacities were about 9,100 kW/km2 and 4,900 kW/km2, respectively, under an assumption of the steam flash system of power generation over 30 years.
At first, we examined what kind of crustal information contributes to increasing the estimation accuracy of geothermal structure by evaluating the feature importance of DNN. The importance of crustal information such as the distribution of active volcanoes, the Curie point depth, rock type, geochemical data of hot springs, temperature distribution at different elevations estimated by the Stacking Method, and geothermal gradient were evaluated as essential factors of the estimation accuracy, in particular, the horizontal temperature distribution, geothermal gradient, and geochemical data of hot springs were the predominant factors.
Next, applicability of NK to all area of Japan Island was examined. The spatial correlation structure of the temperature logging data must be different in each region and between geothermal and non-geothermal regions. Therefore, we constructed a NK model that incorporates the characteristics of the spatial correlation structure in each region. This NK model was proved to increase the extrapolation accuracy and produce smaller error than DNN.
Finally, a three-dimensional geothermal structure analysis was implemented using NK that incorporated all the types of crustal information. A noteworthy result was that the model was able to estimate the geothermal structure from the surface to deep depths over Japan Island from the temperature logging data with limited amount of data points and depth range. In addition, the iso-surface map of the water critical point at 374 °C and 22.1 MPa was constructed by this model, which highlighted promising supercritical geothermal resource areas around Toyoha, Hachimantai, Mt. Kirishima, and Mt. Kurikoma. Furthermore, calculation result of geothermal resource potential revealed that the areas of Toyoha and Mt. Osore in northern Japan are the highest potential in addition to the areas where the geothermal power plants are actually working. Their expected power generation capacities were about 9,100 kW/km2 and 4,900 kW/km2, respectively, under an assumption of the steam flash system of power generation over 30 years.