Real-Time Prediction of Wellbore Trajectory with a Dual-Input GRU(Di-GRU) Model

Abstract Accurate prediction of wellbore trajectory is crucial for precise directional drilling, yet it remains challenging due to the complex underground conditions and the multitude of highly nonlinear factors influencing trajectory variations. Current mechanical analysis methods for trajectory prediction are based on models with numerous assumptions, limiting their simultaneous applicability. Moreover, some existing machine learning algorithms overlook the impact of drilling modes (sliding or rotating) and Bottom Hole Assembly (BHA) types, necessitating improvements in their accuracy. This study developed a Dual-Input GRU Neural Network (Di-GRU) capable of end-to-end prediction, thereby circumventing the need to consider complex underlying mechanisms. The model consists of a time series sub-network and a non-time series subnetwork. The time series sub-network inputs features such as weight of bit (WOB) and rate of penetration (ROP) that vary with time, while the non-time series sub-network inputs features such as geological stratification, BHA type, and drilling mode that do not vary with time. The time series sub-network uses attention mechanism to focus on features that significantly affect wellbore trajectory. To render the non-time series features numerical, one-hot encoding is first used, followed by embedding layer for dimensionality reduction and dense representation. Additionally, a model dynamic update mechanism based on incremental training is established, achieving real-time trajectory prediction and improving the model's adaptability to the complex drilling environment. Data from 12 wells were utilized for the experiments. Through comparative experiments with Gated Recurrent Unit Neural Network (GRU), Fully Connected Neural Network (FCNN), Random Forest (RF) and Support Vector Machine (SVM) models, the results show that the model's prediction accuracy is superior to these three models. Specifically, in terms of the Mean Absolute Error (MAE) of the inclination angle, the model reduces the deviation by 22%, 22%, 58% and 68% compared to the GRU, FCNN, RF and SVM models, respectively; and in terms of azimuth angle, reduces the deviation by 33%, 50%, 84% and 81%. The MAE of the inclination and azimuth angles of the Di-GRU dynamic update model is reduced by 52% and 36%, less than 0.3°, respectively compared to the offline model, indicating that the model has high prediction accuracy and real-time prediction capabilities. This paper presents a real-time intelligent prediction method for wellbore trajectory, taking into consideration drilling mode, geological stratification, BHA structure, and other relevant factors that affect trajectory, such as WOB and ROP. The method has high predictive accuracy and is capable of adapting to changes in complex drilling environments. It also overcomes common problems with current mechanical models such as being complicated, having multiple assumptions which are difficult to simultaneously meet, and intelligent models not specifically considering non-time series features and not being updated in realtime.