<p>With the acceleration of urbanization and the growing demand for underground space development, rectangular pipe jacking has increasingly become an important method in urban underground construction due to its efficient space utilization and adaptability to complex ground conditions. However, the dynamic changes in jacking force during tunnel construction directly impact construction safety and efficiency, and traditional prediction methods have their limitations. To address this issue, this study introduces a jacking force prediction model based on the GRU-KAN algorithm, which combines Gated Recurrent Units (GRU) and Kolmogorov–Arnold Network (KAN). The model enhances prediction accuracy and interpretability through an adaptive activation function. Using field monitoring data from a rectangular pipe-jacking tunnel project in Baoshan District, Shanghai, we constructed a dataset that integrates key features such as jacking distance, frictional resistance, and earth pressure, and trained the GRU-KAN model for prediction. The results indicate that the proposed GRU-KAN model not only substantially improves prediction accuracy but also provides an interpretable prediction mechanism. Specifically, it can reveal the contributions of different input features and capture temporal dependencies, making it particularly valuable for supporting construction decision-making and risk assessment in rectangular pipe-jacking projects.</p>