GEOTHERMAL TEMPERATURE ESTIMATION BASED ON RESISTIVITY DATA USING ARTIFICIAL NEURAL NETWORK: APPLICATION TO THE KAKKONDA GEOTHERMAL FIELD, JAPAN

Abstract

Temperature distribution at a geothermal development provides a vital information for geothermal development. Spichak et al. (2007) has proposed to use artificial neural network estimation to estimate temperature distribution based on resistivity data by Magnetotelluric observation. In this study, we examined the characteristics of the neural network approach at the Kakkonda Geothermal Field in Japan. Temperature data measured at wells were used, including the deepest well of WD1 with the depth of about 3700 m penetrating granite basement rock. Through the application, we showed the importance of the 2D location of teaching data. Specifically, the error increases when teaching data were far from target area. By optimizing the structure of a neural network, we demonstrated that the error of the estimated WD1 temperature was 16 %, and the estimated temperature pattern almost agree with the observed temperature. In addition, we proposed another temperature estimation approach based on neural kriging. Since the neural kriging accounts for variogram of temperature data, underlying statistical structure reflects estimated temperature. As a result of validation, we showed that the estimated temperature successfully reduced the error of data variogram. This result demonstrates the effectiveness of the neural network and neural kriging approach with temperature for estimating temperature at deeper parts.