Estimating temperature distributions in geothermal areas using a neuronet approach

A method is proposed for predicting the distribution of temperatures in geothermal areas using the neuronet approach and, in particular, downhole temperature logs. The method was tested against the results of an analytical model, showing that the errors in neuronet temperature estimates based on well log data derive from: (a) the neuronet “education level” (which depends on the amount and structure of information used for teaching) and (b) the distance of the point at which the estimate is made from the area for which data are available. These conclusions were confirmed when estimating temperatures in eight actual wells, using 50 downhole temperature logs from other wells in the geothermal area. It was found that, for this particular case, neuronet teaching utilizing 30 well logs results in an average forecast error of 20%. As the number of training logs increases (up to 50), the error slightly decreases (down to 16.9%). The effects of the teaching data pattern (conductive-type versus convective-type of temperature profiles) were also studied, and an optimal strategy was developed for the neuronet training, based on the information available.