Real-Time Wellhead Pressure Prediction: An Integration of Deep Learning and Physical Modeling

Abstract

Abstract Hydraulic fracturing, essential for shale reservoir exploitation, relies on wellhead pressure as a key indicator for monitoring and diagnosing fracturing stimulation. The strong heterogeneity of shale reservoirs complicates fracture propagation and proppant transport, leading to complex pressure fluctuations. Accurate modeling and prediction of wellhead pressure during hydraulic fracturing aid engineers in real-time monitoring and risk assessment, providing a basis for adjusting pumping parameters, ensuring a smooth completion of the fracturing plan, and successful production enhancement. Compared to purely data-driven models, we developed a wellhead pressure prediction model integrating deep learning and physical models, considering the actual physical processes of hydraulic fracturing. Appropriate neural network algorithms were selected for various scenarios of fracturing fluid flow, effectively modeling and predicting complex pressure fluctuations in the wellbore and formation fracture systems. The integrated model incorporates a physical model for hydrostatic pressure fluctuations and a parallel (MLP-LSTM) neural network for friction and net pressure changes. We collected over 1391 historical treatment curve datasets from the shale gas fracturing for training and testing. Compared to pure data-driven models, the data-physics integrated wellhead pressure prediction model captures future wellhead pressure trends(increases/decreases) triggered by changes in the pumping procedure and parameters more effectively through the inclusion of a physical model. The optimized data input and the introduction of a parallel neural network structure enable the integrated model to accurately model and learn net pressure fluctuations and friction changes within the formation fracture system, demonstrating the effectiveness of deep learning methods in modeling complex non-linear physical processes such as fracture propagation and proppant transport. The integrated model has been deployed and verified at 18 fracturing stages on the shale gas fracturing site in Wei Yuan, Sichuan Basin, and has successfully predicted rapid pressure increases, alerting field personnel to avoid fracturing screen-out risks 5 times. Based on field test data, the integrated model achieved a wellhead pressure prediction 90 seconds in advance with an RMSPE of 0.07993 and a MAPE of 0.078708. The integrated model accurately predicts pressure trends during treatment, enabling real-time monitoring and adjustment of proppant concentration and flow rate in the main fracturing stage, thereby enhancing fracturing completion rates and production. This research presents the data-physics-driven deep learning approach for wellhead pressure prediction during hydraulic fracturing. By integrating physical models, we enhance model reliability while leveraging deep learning's superior capabilities in learning and modeling complex physical processes. This allows for accurate, efficient wellhead pressure prediction during fracturing treatments, providing engineers with a basis for monitoring hydraulic fracturing, reducing operational complexity, and ensuring effective fracturing stimulation.