To overcome the limitations of conventional sand-plug identification methods during hydraulic fracturing operation, such as low efficiency, high labor-intensity, limited accuracy, and inability to provide real-time early warning, we develop an automated sand-plugging risk identification and intelligent early-warning model for offshore fracturing wells based on multi-parameter operational data—including operational pressure, pumping rate and sand concentration—and deep learning algorithms. Firstly, an attention-based long short-term memory neural network (Att-LSTM) is employed to establish a real-time wellhead pressure prediction model, which could forecast pressure evolution 40 s in advance with an accuracy exceeding 92%. Secondly, an improved attention-based convolutional neural network–LSTM (Att-CNN-LSTM) model is proposed to identify sand-plug, achieving a temporal identification error of less than 1 min. By integrating these two models and incorporating a transfer learning module, a real-time sand-plugging risk early-warning system with continuous transfer learning capability is established. The results indicate that the proposed warning model, driven by the predicted pressure values, can identify sand-plugging events and outputs the sand-plugging probabilities for both the current moment and the subsequent 40 s, calculated as the average of the top five probability values. Field validation shows that the system can trigger warnings 38–42 s prior to actual sand-plugging events. In addition, the embedded transfer learning module helps reduce the number of training iterations required for formal model convergence from 2 000 to 300, improving computational efficiency by a factor of 5.7. This study demonstrates that the proposed deep learning approach can significantly enhance the accuracy and efficiency of sand-plug identification and early-warning, thereby accelerating the intelligent decision-making process in hydraulic fracturing operations.
Guo et al. (Thu,) studied this question.