Prediction method for formation pore pressure based on transfer learning

Abstract

Pore pressure prediction of drilled wells is the recognition of formation information using logging and other actual drilling data after drilling is completed, which is of great significance in improving the mastery of regional formation information and reducing the potential risks of the engineering program of wells to be drilled. Traditional formation pressure prediction methods rely on manual experience and are difficult to cope with complex formations. Existing machine learning-based methods have high accuracy in the prediction of formation pressure for wells with similar data composition in the same block or logging, but have poor accuracy in the prediction of wells with large differences in data from different blocks or logging. In this paper, a transfer learning algorithm is introduced to modify the existing LSTM-based formation pore pressure prediction model and establish a transfer learning-based formation pore pressure prediction method. In addition, we evaluated the prediction performance of machine learning algorithms and transfer learning algorithms under the same/different blocks and the same/different logging data conditions. The results show that, compared with the existing machine learning algorithms, for the same block and the same logging data conditions, transfer learning can improve the prediction accuracy by 0.15%; for the same block and different logging data conditions, transfer learning can improve the prediction accuracy by 0.33%; for the different block and the same logging data conditions, transfer learning can improve the prediction accuracy by 0.19%; for the different block and different logging data conditions, transfer learning can improve the prediction accuracy by 0.11%. Overall, the addition of the transfer learning method can improve the accuracy and generalization ability of the formation pore pressure prediction model based on machine learning to a larger extent.