Data-Driven Prediction of Carbonate Formation Pressure Using LSTM-Based Machine Learning

Deep to ultra-deep carbonate formations have become crucial targets for oil and gas exploration. However, owing to the low accuracy of carbonate formation pressure prediction during drilling, complex incidents such as collapse, block shedding, and drilling fluid loss frequently occur, severely restricting the efficient development of deep and ultra-deep oil and gas resources. This study targets the Tarim Basin, integrating well-logging and geological data from six wells, with depths ranging from 5000 to 9000 m, through multi-source data fusion. These results indicate that abnormal overpressure in the carbonate formations is chiefly governed by hydrocarbon generation and tectonic compression. Accordingly, 10 key characteristic parameters related to the cause of over-pressure were identified. The Support Vector Regression (SVR) model and Long Short-Term Memory (LSTM) neural network model were used to predict the pressure of carbonate rock formations. The constructed LSTM model demonstrated better prediction results for formation pressure than the SVR model. Compared with the traditional Bowers effective stress method, the LSTM model achieves an exact mean relative error range of 0.256–3.846% for a single well, which is significantly lower than the prediction accuracy of the Bowers effective stress method. The study shows that the LSTM machine learning algorithm can more accurately predict the formation pressure distribution characteristics of the carbonate formations in the research area. This provides reliable foundational data support for safe drilling in the carbonate rock formations of the Tarim Basin and offers valuable insights for pressure prediction in similar regions.