Spatial-Augmented XGBoost for Railway Track Geometry Forecasting under Limited Temporal Observations

Abstract

Track geometry forecasting under limited temporal observations represents a persistent challenge in railway infrastructure management, particularly in networks where systematic inspection programs are recent and historical records span only a small number of cycles. Precise track geometry is fundamental to ensuring operational safety, passenger comfort, and the dynamic stability of the train as it traverses the rail network. This paper proposes Spatial-Augmented XGBoost (SA-XGBoost), a framework that compensates for shallow temporal depth by systematically encoding spatial neighborhood information as model input. Five feature groups are constructed from inspection snapshots: multi-hop neighbor values, spatial gradients, rolling statistics, temporal degradation dynamics, and missingness indicators. The model targets incremental geometry change rather than absolute values, with predicted increments constrained by physically derived bounds. Evaluated on 219,000 measurement points from the Jakarta–Bogor railway corridor across seven geometry parameters, the framework achieves strong performance on cant and alignment parameters (R2 = 0.95 and 0.65–0.66 respectively) while performance on longitudinal profile and twist parameters remains limited, reflecting the higher spatial irregularity of those degradation patterns. These results demonstrate that spatial augmentation is a viable strategy for geometry forecasting under data-scarce conditions, with parameter-dependent effectiveness that warrants further investigation.