Control-risk-informed spectrum-power allocation for virtual coupling of heavy-haul trains

Heavy-haul train virtual coupling requires spectrum-power allocation that protects both packet-level quality of service (QoS) and the freshness of train-control commands. This paper proposes CRISP, a control-risk-informed spectrum-power allocation method implemented on a low-band OFDMA resource-grid abstraction. The method combines a double deep Q-network with TCS-prioritized replay, feasible-action filtering, and guard-band fallback. A timestamped command-age state is propagated across packet, resource-allocation, and train-control clocks, and a physics-calibrated communication-induced control-degradation index combines spacing error, velocity error, and command staleness. Under the reference tunnel stress condition (four trains, four schedulable RBs, adjacent-channel leakage coefficient 0.1), CRISP reduces the TCS QoS violation rate to 0.8085 (SD 0.0112), compared with 0.9912 (SD 0.0111) for Vanilla-DDQN and 0.8548 (SD 0.0170) for Guard-Band-Greedy. Mean TCS command age is reduced to 94.0 ms from 1566.4 ms and 377.8 ms, respectively; no degradation event is observed and the mean minimum braking margin remains 987.5 m. Ablation, learning-convergence, decision-latency, and emergency-outage experiments identify the contribution and boundary of each module. Because the reference condition is intentionally communication-stressed, the results demonstrate degradation mitigation rather than field reliability certification or a formal safety guarantee.