CSI-Aging-Aware Functional Split Optimization and KAN-LSTM Prediction for Open-RAN NTN

Abstract

In this paper, we study functional-split selection for an Open radio access network (RAN) deployed as a non-terrestrial network (NTN), in which a low-Earth-orbit (LEO) satellite hosts the radio unit, so that each candidate split (7.2x, 6, 2) places the RAN at a distinct point of the transparent-to-regenerative axis. We capture the space-ground geometry exactly through an instantaneous link plane, which yields closed-form slant range, elevation, delay, and Doppler for an inclined orbit and a rotating Earth. On this geometry we derive the received-SNR density of the elevation-driven LOS/NLOS mixture channel and, through a split-dependent channel state information (CSI) aging mechanism, an uncoded bit-error-rate (BER) law for each split, establishing that what separates the splits is the freshness of the CSI each brings to its adaptation loops: a loop closed on the ground pays the feeder round-trip and incurs an aging-induced error floor. We formulate split selection as a prefer-centralized feasibility problem over reliability, aging, and fronthaul-capacity constraints, and we show that its exact solution is a closed-form region rule in four state coordinates, validated against exhaustive enumeration without a single mismatch. Because this solution is reactive whereas a split change must be initiated ahead of time, we propose a compact Kolmogorov-Arnold network (KAN)-LSTM, sized for an edge xApp/rApp, that predicts the split schedule seconds ahead; we show that it attains high schedule accuracy at a sub-percent BER-violation probability and delivers nearly all of the ground-truth policy's goodput, outperforming naive and reactive baselines.