A cost-efficient data-driven framework for rail temperature prediction and thermal buckling risk assessment

Abstract High rail temperatures compromise the integrity of continuously welded rail (CWR), generating compressive forces that may trigger track buckling. This study proposes a cost-effective, data-driven framework for rail temperature prediction using only two measured inputs, ambient temperature and relative humidity, both routinely measured by standard meteorological stations, together with temporal features obtained from the timestamp. No solar radiation, wind or cloud cover sensor is required, thus avoiding complex sensor arrays. The prediction is converted into the axial thermal stress governing buckling, the quantity used in track engineering. Five machine learning algorithms were benchmarked: Artificial Neural Networks (ANN), Support Vector Regression (SVR), Random Forest (RF), Long Short-Term Memory networks (LSTM), and XGBoost. Two strategies were evaluated: a long-term global model on a two-year dataset, and a short-term sliding-window model predicting up to three hours ahead. Models were trained on measurements from an active Brazilian freight railway (2022–2023) and validated on the temporally separate January to May 2024 window, the period of maximum buckling exposure. Hyperparameters were optimized via Optuna using a Tree-structured Parzen Estimator. ANN delivered the lowest long-term error ($R^{2}$ = 0.946; MAE = 1.54 $^{\circ }$C), followed closely by SVR ($R^{2}$ = 0.941; MAE = 1.63 $^{\circ }$C), while SVR achieved the best short-term precision at the one-hour horizon ($R^{2}$ = 0.949; MAE = 1.09 $^{\circ }$C, with RF closely following at MAE = 1.10 $^{\circ }$C). Expressed as rail axial stress, these errors correspond to about 3.67 and 2.59 MPa, roughly 7.7% and 5.5% of the compression developed by a 20 $^{\circ }$C excess over the neutral temperature. Despite the limited inputs, the framework achieves competitive accuracy, a practical solution for sparsely monitored networks.