Locating a New Drainage Well by Optimization of a Back Propagation Model

Abstract

Geoelectric (GE) field data are usually interpreted using GE computer-codes. A back propagation (BP) ANN model is used to predict hydro-sedimentology (HS) well logs. Both twenty-six sets of existing well logs from Golgohar iron mine and data from 26 Schlumberger GE apparent resistivity surveys were used as inputs for the ANN model. This data file was used to expend ANN models during the learning and testing phases of the modeling and the results were compared with the HS logs of the drilled wells. The learning phase file was included 20 records and the testing phase 6. The Professional II/Plus computer-code was used to make the various ANN models. In this stage, two techniques were employed: (1) incorporating hydro-sedimentology (HS) codes, (2) switching the GE field-observed scatter diagram to a digit string. None of the model results were able to satisfy the authors’ expectations, possibly due to inadequate input data, so the model was optimized in second stage using the third novel technique that could complete it: using both field data and interpreted GE computer-code (IPI2Win) results as appended input data. The best BP-ANN model had three hidden layers, and a DBD learning rule, sigmoid transfer function, and epoch learning of 15,000,000 with 0.000007 Root Mean Square (RMS) error and provided better conformity between predicted and observed HS logs. Employing 3 mentioned techniques was useful for locating a new drainage well at the Golgohar mine.