A Sequence Learning Approach for Real-Time and Ahead-of-Bit Pore Pressure Prediction Utilizing Drilling Data from the Drilled Section

Summary Accurate pore pressure prediction is vital for ensuring drilling safety and efficiency. Existing methods primarily rely on interpreting logging while drilling (LWD) data for real-time pore pressure prediction. However, LWD tools typically collect data from sensors located approximately 100 ft behind the drill bit, reflecting formations that have already been penetrated rather than those being actively drilled. In contrast, drilling data reflect the formations being actively drilled at the bit, without requiring additional downhole equipment or extra costs. Nevertheless, traditional methods using drilling data often employ simplified theoretical formulas that oversimplify the complex characteristics of geological conditions. Although a few studies have utilized machine learning with drilling data for prediction, they often employ point-to-point prediction methods, using drilling data from a given depth to predict pore pressure at the same depth. This approach overlooks the sequential nature of geological formations along the well depth, limiting prediction accuracy and the ability to forecast pore pressure ahead of the bit, which is crucial for proactive decision-making. Therefore, this study proposed a novel approach that utilizes historical drilling data from the upper drilled section (drilled window) to predict pore pressure, specifically employing two prediction methods: (1) Real-time predictions use a sequence-to-point strategy, where drilled window data are used to predict pore pressure at the drill bit. (2) Ahead-of-bit predictions employ a sequence-to-sequence strategy, where drilled window data are utilized to predict undrilled formations ahead of the drill bit. This study developed three custom-designed neural network models using long short-term memory (LSTM) and self-attention algorithms: LSTM, Double-Layer LSTM, and LSTM-Attention. For real-time prediction, a Double-Layer LSTM model with a 15-m drilled window length achieves stable performance with a mean squared error (MSE) of 1.45×10⁻⁴. Integrating drill bit characteristics further improves accuracy, increasing the coefficient of determination (R²) from 0.61 to 0.89 for Well Test-1 and from 0.50 to 0.75 for Well Test-2. Field tests on two ongoing drilling wells demonstrate the practicality and robustness of the proposed approach, achieving R² values of 0.72 and 0.83. For ahead-of-bit prediction, the study provides reference guidance for pore pressure prediction at distances of 10, 20, 30, and 40 m ahead of the bit, presenting optimal configurations for each scenario. The LSTM-Attention model demonstrates superior performance. However, as the prediction distance increases, the prediction error also grows. The recommended configuration for this data set is a prediction distance of 30 m ahead of the bit with a drilled window length of 80 m, yielding an MSE of 2.88×10⁻⁴. This configuration strikes a balance between prediction accuracy and prediction distance, ensuring the maximum prediction distance while maintaining an acceptable level of accuracy. Field operators can flexibly choose the prediction configuration based on their specific requirements for accuracy and prediction distance. This study could achieve accurate real-time and ahead-of-bit pore pressure predictions, facilitating the early identification of drilling risks and enabling timely adjustments, thereby improving drilling safety and efficiency.