A spatiotemporal relative-ring modeling framework for construction-stage ground settlement prediction in shield tunnelling

Construction-stage ground settlement prediction in shield tunnelling remains challenging because settlement monitoring is sparse and irregular, whereas shield operation data are dense and closely linked to the moving construction process. Existing time-based forecasting and direct regression approaches often compress the effects of successive construction stages into a current or terminal response and therefore cannot explicitly represent settlement evolution during shield passage. This study proposes a spatiotemporal relative-ring modeling framework that reorganizes tunnelling data according to the relative position between the shield and each monitored section. The framework predicts stage-wise settlement increments and reconstructs the cumulative settlement response within a section-centered relative-ring window. To accommodate sparse supervision, a bin-aggregated mask-aware multi-task model is developed to jointly learn process-level settlement increments and the window-end cumulative response. The framework is evaluated using field data from Guangzhou Metro Line 18 under subsequent-section and cross-line validation protocols. Results show reconstructed window-end MAEs of 0.84 mm and 0.79 mm, respectively, outperforming conventional local-feature regression and sequence-based baselines. The results indicate that relative-ring organization, bin aggregation, and process-oriented supervision provide an effective and interpretable approach for construction-stage settlement prediction under sparse field monitoring.