A Novel Hybrid Transfer Learning Method for Bottom Hole Pressure Prediction

Abstract

Abstract Accurate and efficient prediction of bottom hole pressure (BHP) is critical for safe drilling in complex formations. Recently, more and more studies have found that intelligent models have high accuracy in predicting BHP. However, most of the existing intelligent methods are limited to historical measurement, which is insufficient in processing new data. To overcome the problem of insufficient generalization ability of the model on new dataset, we propose a novel hybrid transfer learning approach to incrementally predict BHP in different well sections, combining long short-term memory (LSTM) and domain adversarial neural network (DANN). For deeper well sections, this method applies features learned from historical data in the upper well sections to boost the BHP prediction. LSTM is applied to extract temporal characteristics from the upper and deeper well sections. DANN tries to discover constant characteristics between the upper and deeper well sections. Next, LSTM-DANN model trained with data in the upper well sections can be used to support prediction of target BHP without degradation of accuracy due to data offset. Based on drilling data from the field, the performance of the proposed method is fully assessed. Results indicate that the hybrid transfer method can significantly improve the BHP prediction performance compared to models trained on data from different well sections. The mean absolute percentage error of the novel method reaches 0.15%, which is reduced by 30% compared with the original one. This study provides reference for accurate managed pressure drilling, and contributes to improve the transferability and generalization of intelligent models.