Geothermal is a clean and abundant energy resources. For an assessment of geothermal resources, clarification of the temperature distribution and hydrothermal systems in deep parts around hot springs and fumarole manifestations is indispensable. The most reliable data for such an assessment can be obtained by temperature logging. In general, the distribution of geothermal wells is biased, and their temperature data are limited in depth ranges. Therefore, the estimation of temperature from the surface to deep zones is difficult from a well-logging data set. To overcome this problem, we examined a combination of neural networks and geostatistics. A feedforward neural network was used to extrapolate temperature logging data of each geothermal well, and the temperatures estimated at the wells were interpolated using geostatistics. The 22 km × 18 km region in the Hohi geothermal area in central Kyushu, southwest Japan, was chosen as the test site, and temperature data from 20 wells were used in the analyses. The wells were classified into two types, conductive (13 wells) and convective (7), based on their patterns of temperature change with depth. Two networks for conductive and convective type data were trained separately. Binary values, 0 and 1, were assigned to the well locations depending on conductive or convective type. A semivariogram was constructed from the binary data and used in ordinary kriging so that the spatial correlation structure concerning the pattern of vertical changes of temperature can be reflected in the interpolation. It was clarified that the extrapolation of temperature data from the lowest level in a well down to -2,000 m can be appropriately performed at each location in spite of any shortness of the measured depth range. Accordingly, three-dimensional temperature distributions in the study area could be characterized through the proposed method.
View full abstract