Counterfactual Root-Cause Analysis for Traction-Power Disturbances in High-Speed Railways

Abstract

High-speed railway traction power systems rely on continuous monitoring of traction current, catenary voltage, transformer temperature, feeder load, breaker status, harmonic distortion, return current, and protection relay signals. These variables are causally linked through power conversion, train operation schedules, load transfer, and protection-control mechanisms. Abnormal events may appear first as small electrical deviations and then propagate across substations, feeders, and onboard traction equipment. This study proposes a counterfactual root-cause analysis method for multivariate traction power system anomalies in high-speed railways. The method learns a time-lagged causal dependency structure among electrical, thermal, and protection variables under different operating regimes. For each abnormal segment, counterfactual trajectories are generated by intervening on suspected driver variables, and root-cause ranking is performed according to the reduction in downstream residual deviation. Experiments are conducted on a traction power monitoring dataset covering 62 substations, 184 feeder sections, 416 protection relays, and 73 monitoring variables collected at 2-second intervals over 13 months. The dataset contains 516 million time-stamped records and 1,540 maintenance-verified abnormal events, including feeder overload, catenary voltage sag, transformer thermal rise, harmonic disturbance, breaker misoperation, and relay-delay events. The proposed method shortens median root-cause analysis time from 29.3 minutes to 8.5 minutes compared with a temporal graph reconstruction model. The average fault-section localization error decreases from 4.1 feeder sections to 1.3 feeder sections. Causal intervention analysis identifies primary driver variables in 1,170 events before protection signals become dominant. The median online inference latency is 34 ms per monitoring window, supporting near-real-time railway operation monitoring. These findings indicate that counterfactual causal analysis can provide fine-grained and interpretable anomaly diagnosis for high-speed railway traction power systems.

Keywords: High-speed railway; traction power system; causal inference; counterfactual analysis; multivariate time series; root-cause diagnosis; power system anomaly detection.