<p>Accurate prediction of fracturing pressure is crucial for evaluating the morphology of hydraulic fractures and achieving early warning of proppant plugging. The strong heterogeneity of reservoirs and the complex fluid‒solid coupling mechanism of fracture propagation make physical modeling for predicting fracturing pressure time-consuming and labor-intensive, with poor generalizability. Consequently, the real-time accuracy requirements are difficult to meet in practice. In this study, a data-driven dual-layer network model is proposed to address the real-time prediction of fracturing pressure and the optimization of proppant plug prevention parameters. The upper-layer network uses a CNN-LSTM-Attention model to predict the fracturing pressure–time series, whereas the lower-layer network employs a LightGBM model to address the issue of comprehensive spatial feature effects and improve the generalization capability. Furthermore, an ensemble learning algorithm based on error correction is introduced to adjust the weights of the dual-layer network structure dynamically and adaptively across different time steps. The application results show that the established dual-layer network model can simultaneously capture the long-term temporal dependencies of the global time series and the spatial features affecting fracturing pressure, thus enhancing model accuracy and generalizability. This method accurately predicts pressure anomalies and enables advanced prediction of fracturing pressure. By regulating the pump rate and proppant ratio, which most considerably impact pressure, six different schemes are recommended. Implementation of the recommended scheme ensured, successful execution of the operation.</p>