A data-driven prediction method over the lifecycle of fracturing and production of horizontal wells in shale

Efficient determination of hydraulic fracture lengths and comprehensive comprehension of post-fracturing production performance are crucial for optimizing fracturing design and maximizing ultimate recovery. Nevertheless, numerical simulation methods demand high professional expertise and entail significant computational resources. Meanwhile, the complex coupling of multiple factors makes the direct application of existing neural networks to hydraulic fracturing scenarios challenging. To efficiently forecast the post-fracturing fracture network and the dynamic evolution of the pressure field during production, an intelligent prediction model for post-fracturing fractures is developed by integrating the fast Fourier transform structure, convolutional auto-encoder, and U-Net structure. This model is capable of predicting the intricate fracture morphology after fracturing. Furthermore, leveraging a feature fusion architecture combining the convolutional long short-term memory neural network and the channel attention module, a dynamic prediction model for the pressure field during post-fracturing production is proposed to capture pressure field fluctuations over time. The combination of these two models yields an intelligent prediction approach for horizontal well fracturing. The performance of the intelligent prediction model for post-fracturing fractures is evaluated by means of the F1 score, and a score higher than 0.90 is accomplished. The absolute error is used to assess the dynamic prediction model of the pressure field during post-fracturing production, with a single grid pressure error less than 1.0 MPa. The research findings suggest that the proposed intelligent prediction method for fracturing operations can enhance prediction efficiency and provide novel insight for optimizing fracturing design and augmenting final production capacity.