A data-driven approach for rail temperature estimation from air temperature, solar irradiation, and land surface temperature

• Rail temperature was estimated using multivariate regression and machine learning. • A scenario-based framework was proposed for practical rail temperature estimation. • Linear regression achieved accuracy comparable to ML with higher practicality. Excessive summer heat induces thermal expansion of railway rails, increasing the likelihood of buckling accidents and posing a serious threat to train operation safety. In South Korea, speed restrictions and operational controls are implemented based on Rail Temperature (T R ) thresholds to mitigate such risks. However, direct measurement of T R across the entire railway network is practically difficult. Accordingly, Air Temperature (T a ) has commonly been used in South Korea to estimate T R . Nevertheless, this approach is limited in its applicability to sections where T a is not observed and exhibits inherent limitations in prediction accuracy. To address this limitation, this study proposes a method for estimating T R in sections where direct measurements are unavailable. Various meteorological variables collected from nearby Automated Synoptic Observing System (ASOS) stations were analyzed using SHapley Additive exPlanations (SHAP) to identify key factors influencing T R , and a total of seven scenarios were constructed based on different combinations of three selected variables. For each scenario, both regression based (Linear Regression (LR), RANdom SAmple Consensus (RANSAC), and Theil–Sen) and machine learning based models (Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Networks (ANN)) were applied. Scenario performance was first evaluated using Taylor diagrams, followed by quantitative assessment using Root Mean Squared Error (RMSE), Bias, RMSE–Standard deviation Ratio (RSR), Nash–Sutcliffe model Efficiency coefficient (NSE), Pearson correlation coefficient (Pearson’s r), and the McNemar test to determine the most effective methodology. The results indicated that LR was the best-performing regression based model, while RF achieved the highest performance among machine learning models. Furthermore, the regression equations derived from the proposed scenarios outperformed the conventional T R estimation formula currently used in South Korea. This study is expected to contribute to railway operational safety by enabling reasonably accurate T R prediction in terrains where installing T R sensors is difficult, particularly under high buckling risk conditions.