Research on Prediction of Polymer Porous Grouting Effect in Voided Soils Based on CNN-BiGRU

Abstract To improve the prediction accuracy of the polymer porous grouting effect in voided soils (especially the topsoil response), an intelligent CNN-BiGRU prediction model is constructed, which integrates the spatial feature extraction capability of Convolutional Neural Network (CNN) and the time-series modeling ability of Bidirectional Gated Recurrent Unit (BiGRU). Taking grouting volume and void depth as input parameters, and the soil responses at both borehole ends, one borehole end, and the top as output targets, 35 sets of experimental datasets are built. After normalization, the CNN-BiGRU model is compared with the single CNN model and single BiGRU model. Each model is trained independently 5 times, and its performance is comprehensively evaluated using training time (T), coefficient of determination (R²), mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE). The results show that all three models can effectively predict the grouting effect (average R² >0.9), but the CNN-BiGRU fusion model outperforms the single models and has the highest training efficiency: the average T is 8.3 s (CNN: 8.6 s; BiGRU: 10 s), the average R² reaches 0.9455 (CNN: 0.9403; BiGRU: 0.9167), the average MAE is 0.1007 (CNN: 0.1358; BiGRU: 0.1111), the average RMSE is 0.1413 (CNN: 0.1807; BiGRU: 0.1647), and the average MAPE is 0.0271 (CNN: 0.0611; BiGRU: 0.0338). In the prediction of topsoil response, the CNN-BiGRU fusion model has more advantages, with R² as high as 0.9896 (CNN: 0.9106; BiGRU: 0.8983) and MAE as low as 0.0965 (CNN: 0.2537; BiGRU: 0.1666). The relevant research can provide a new method for predicting the polymer porous grouting effect in voided soils and a reference for optimizing grouting parameters.