Real-Time Prediction of Bottomhole Pressure Based on Data Expansion and Incremental Update

The data of complex well sections are difficult to acquire and the static prediction model is inapplicable to the fluctuating downhole conditions.Considering these challenges, the intelligent method was proposed to predict the fluctuating bottomhole pressure(BHP)in a real-time manner.Based on the generative adversarial network(GAN), which helps to expand the downhole single data measured while drilling and solve the problem of limited effective BHP data, the BHP prediction model with data enhancement was established.In order to effectively improve the adaptability and transfer performance of the model to the variable working conditions, under the condition of incremental update of data, multiple long short-term memory(LSTM)models were trained in a segment-wise manner, and the real-time update of the prediction model was delivered on the basis of transfer and ensemble learning.Finally, the hybrid attention mechanism was used to realize the interpretable analysis of the intelligent prediction model.The results show that the model after data expansion training presents a significant superiority in accuracy and stability, and the incremental update real-time prediction method greatly improves the generalization capacity of the model and the time-effectiveness in field applications.The average relative error of the model is only 0.12%.In addition, the interpretable analysis shows that the BHP has considerable short-term autocorrelation and the wellhead back pressure is characterized by fluctuating transfer.The research findings provide a theoretical guidance for accurate and efficient prediction of BHP and interpretability of intelligent models for drilling deep wells.