Logical analysis of survival data for remaining driving range prediction in battery electric bus operations

The remaining driving range (RDR) of a battery electric bus (BEB) is the predicted distance that the bus can travel before its battery reaches its operational cut-off state of charge (SOC). This paper proposes an interpretable probabilistic model that combines Logical Analysis of Data (LAD) with Kaplan–Meier (KM) survival estimation to model the distance to cut-off SOC as a survival variable and obtain probabilistic RDR predictions. LAD extracts interpretable patterns that link the operating conditions to range classes defined based on SOC, while KM estimates nonparametric survival curves that capture the statistical distribution of the distance to cut-off SOC. The LAD–KM model is designed to operate when historical runs include both complete and right-censored runs. Right-censoring, which is common in fleet operations, is when a service run ends before the SOC reaches the cut-off level. This can be the result of scheduled withdrawal, intermediate charging, operational interruptions, or data-logging limits. In these cases, the true distance to cut-off is known only to exceed the observed distance. The model developed operates at distance checkpoints, in which recent driving and environmental conditions are summarised over 1 km and 5 km distance windows. At each checkpoint, SOC values are clustered using k -means to define range classes. LAD then learns patterns that discriminate between these classes and a pattern-specific KM curve is estimated from the historical data of BEB runs covered by each pattern. For a new run and at each checkpoint, survival information from all matched patterns is aggregated to predict the distance to cut-off SOC and the corresponding RDR. The model is evaluated using uncensored real-world BEB runs from urban and suburban routes in Beijing during winter, spring/autumn, and summer season conditions. Using the 0.3 quantile of the conditional survival distribution of the distance to cut-off SOC, selected based on validation performance to favour conservative predictions, LAD–KM achieves a mean absolute error of 7.0 km, a mean absolute percentage error of 3.2%, and a predicted-before-actual-failure rate of 79%, with uncertainty quantified through run-level bootstrap confidence intervals. These results indicate that the model achieves predictive accuracy while limiting late predictions, which correspond to battery depletion before the route is completed. The proposed model, therefore, provides an interpretable decision-support tool for BEB dispatch and fleet-management frameworks in BEB operations, although broader validation on larger multi-vehicle and censored datasets is still necessary.