Physics-Grounded Deep Reinforcement Learning for Power Control in LEO Non-Terrestrial O-RAN: A QoS-Power Pareto and Cross-Orbit Robustness Study

Abstract

We ask when learning helps for O-RAN non-terrestrial power control, and answer with a leakage-free fair-baseline protocol on a causal SGP4+3GPP-TR-38.811 LEO-NTN uplink environment. A converged PPO policy is measured, over 35 disjoint held-out SGP4 passes (hierarchical cluster bootstrap, Holm correction), against a fair classical family: open-loop fractional power control (OLPC), adaptive-PRB greedy control, and a TS 38.213 closed-loop TPC; and a perfect-model certainty-equivalence oracle (Kalman fade prediction + chance-constrained minimum power). Across the full operating envelope (two link margins x two HARQ depths), PPO is the best controller nowhere: a classical controller significantly exceeds it at three of the four corners and ties it at the fourth. The envelope has structure. At moderate margin (G/T = 6.5 dB/K) the perfect-model oracle is strongest, dominating PPO by up to D = +0.063 SLA satisfaction at half its transmit power, with the deployable greedy heuristic close behind (17-35% less power than PPO). Under severe starvation (G/T = 1.1 dB/K) the deployable greedy heuristic is best, dominating PPO with HARQ off and matching it at NR-default HARQ, where the starved regime levels every controller. This is the certainty-equivalence prediction for a known-model, correlated-fade problem: here, a converged model-free policy (PPO) does not beat a well-tuned formula. We contribute the causal, O-RAN-integrable environment and fair protocol as a reusable benchmark, an E2SM-NTN E2 service model exposing NTN link-budget state and a power-control action, and an honest map that scopes learning's value to the multi-cell-interference and bursty/URLLC regimes where no closed-form optimum exists.