Predicting Rate of Penetration of Horizontal Wells Based on the Di-GRU Model

Horizontal wells have become the predominant well type for developing complex oil and gas reservoirs, yielding favorable development outcomes. Accurately predicting the rate of penetration (ROP) is of great significance for cost reduction and efficiency improvement in the drilling process of horizontal wells. Owing to the continuous demand for directional drilling and the intricate trajectory, it is prone to induce complex mechanical behaviors, such as pipe buckling and wellbore stability. It is difficult to accurately predict ROP of horizontal wells. To solve this problem, a novel data-driven ROP prediction model for horizontal wells is proposed by combining a self-attention-based Gated recurrent unit (GRU) network and Back-propagation (BP) Neural Network. At first, the selection of the model's processed input characteristics, in addition to the azimuth and inclination of the well trajectory, is guided by physical laws and statistical analyses. Then, temporal features, such as weight on bit and rotary torque, are input into GRU network, and non-temporal features, such as bottom hole assembly (BHA) and formation layers, are characterized and dimensionally reduced through unique hot encoding and embedding layers before being input into the BP neural network. Next, to accomplish real-time prediction, the model is dynamically updated utilizing the sliding window approach and data flow after training with various well section (vertical, build-up and horizontal well) data. Finally, the model's performance is comprehensively analyzed using field drilling datasets after optimizing model hyperparameters through the orthogonal experiment method. Compared with the BP and LSTM models, the mean absolute percentage error is separately reduced 9.09% and 5.99%. By introducing an incremental update mechanism, the prediction mean absolute percent error (MAPE) of the dynamic update model is reduced by 5.42% compared to the offline model. Results indicate that the model demonstrates strong superiority due to the self-attention-based dual-input network structure, segmented modeling, and dynamic update mechanism. The model exhibits high accuracy and robustness, offering valuable theoretical support for the optimal and efficient drilling of horizontal wells.