A New NARX Neural Network Model for Real-Time Pore Pressure Prediction Based on Surface Logging Data

Pore pressure (Pp) is crucial for oil and gas exploration and development, and accurate prediction ensures safe and efficient drilling. In this paper, a novel method for real-time Pp prediction is proposed based on the NARX neural network method by considering variation patterns of surface logging data. This approach uses upper formation data to predict the lower one. We collected surface logging data from two wells at South China Sea. The surface logging data of one well was used for modeling, while the other serves for an additional test to evaluate the generalization performance of the method. Six surface logging parameters, including hook load (WOH), weight on bit (WOB), torque (TOR), flow rate (FLW), rate of penetration (ROP), and stand-pipe pressure (SPP), were selected as inputs for the model, with Pp gradient as the output variable. Similar to the time delay window in the NARX neural network model, this paper introduces a sequential length to include proper information of upper formations. The optimal sequential length is 10 m, resulting in the best-fitting model with a determination coefficient (R2) of 0.9918, a root mean square error (RMSE) of 0.00459 g/cm3, and a computational time of 18.34 s. Furthermore, this selected model is successfully applied to neighboring wells within the same block, achieving notably high prediction accuracy.