Jacking Force Prediction for Long-Distance Pipe by Integrating Physical Information and Adversarial Learning Mechanism

Accurate prediction of jacking force is crucial for long-distance pipe jacking construction. This study establishes a theoretical jacking force model considering the soil pressure arch effect, where the Persson contact model and the hydrodynamic parallel plate model are incorporated to accurately characterize pipe–soil friction and lubrication resistance, resulting in more reliable friction force estimation as confirmed by field measurements. To enhance prediction accuracy, a Physics-Informed Domain Adversarial Training Neural Network (PIDANN) is proposed, integrating physical constraints into neural network training and expanding the dataset with theoretical values. A case study on the Zheng-Kai city water supply project demonstrates that (1) the theoretical model improves friction force predictions by incorporating the pressure arch effect; (2) PIDANN achieves superior prediction accuracy compared to models without adversarial training; (3) the two fusion methods of physical information integration reduce the mean squared error (MSE) by 36.9% and 20.2%, enhancing generalization and reducing overfitting risks. These findings provide valuable guidance for jacking force control in pipe jacking construction.