PaperlayerView exact revision on arXiv

Source-linked AI summary

StarRTT: An End-to-End Real-Time Framework for Blind Zone-Free RTT Prediction in Starlink LEO Satellite Networks

Jaeyong Lee, Seongjin Jeong, Yusang Park, Hyunho Hwang, Sangmin Han, Wonjae Shin

arXiv:2610.32540v1cs.ITcs.AImath.IT

TL;DR

Serving-satellite changes make LEO RTT non-stationary, while cycle-based identification leaves a blind zone in the information needed for real-time prediction. StarRTT combines per-second identification with RTT prediction, and RT-LSTM achieves 1.405 ms average MAE while reducing error against LSTM, StarNet, and ASTM.

Abstract

from arXiv · show

Low Earth Orbit (LEO) satellite mega-constellations such as Starlink require frequent handovers due to high orbital velocities, resulting in substantial Round-Trip Time (RTT) fluctuations. Accurate RTT prediction can mitigate these fluctuations, yet it fundamentally relies on continuously updated serving- satellite information. Existing serving-satellite identification methods typically complete identification only at the end of each scheduling cycle, creating an identification blind zone that limits real-time prediction. To overcome this limitation, we develop StarRTT, an end-to-end real-time framework that eliminates the blind zone by integrating per-second serving-satellite identification with RTT prediction. Within this framework, we propose Real-Time Long Short-Term Memory (RT-LSTM), an LSTM-based RTT prediction model designed to capture both non-stationary RTT evolution and time-varying geometric relationships among the serving satellite, user terminal, and gateways. Specifically, a historical LSTM encodes historical RTT observations and real-time serving-satellite and gateway-relative features, while a future LSTM processes Two-Line Element (TLE)-propagated future features of the identified serving satellite up to the next scheduling instant. Under strict real-time back-testing, RT-LSTM achieves a Mean Absolute Error (MAE) of 1.405 ms averaged over all 15 prediction horizons and 15 reference time-slot indices, reducing the error by 4.3 %, 11.6 %, and 22.1 % compared with the existing baseline models, LSTM, StarNet, and Attention-based Spatiotemporal Model (ASTM), respectively. Notably, the reduction over LSTM reaches 12.5 % at the scheduling instant, highlighting the benefit of real-time serving-satellite information when recent RTT observations are limited.

I. INTRODUCTION

Frequent handovers make LEO RTT unstable, while existing cycle-based serving-satellite identification leaves gaps in the current information needed for real-time prediction. StarRTT addresses this gap by pairing per-second identification with RTT prediction.

I. INTRODUCTION

  • Starlink satellites travel at approximately 7.5 km/s, and the 15-second scheduling cycle accompanies an average single-satellite connection duration of approximately 28.5 seconds.Frequent handovers are identified as a major source of latency instability in LEO satellite networks.Source: §I · p. 2

satellite along the communication path, RTT exhibits abrupt fluctuations in LSNs. Accurate

  • Serving-satellite changes make RTT non-stationary, and operators’ undisclosed serving-satellite information must be inferred from UT observations and TLE-predicted trajectories.The paper identifies handovers as a source of latency instability in LEO satellite networks.Source: satellite along the communication… · p. 2, §I · p. 2

second window against TLE-based candidate trajectories using the Dynamic Time Warping

  • Cycle-based identification leaves an identification blind zone of up to 14 seconds, during which prediction systems may use stale serving-satellite information.The limitation is temporal resolution: even an accurate cycle-end result may no longer be valid if the serving satellite changes.Source: second window against TLE-based… · pp. 3–4
  • StarRTT jointly performs per-second serving-satellite identification and multi-step RTT prediction to eliminate the blind zone in scheduling-cycle-based approaches.The proposed framework updates serving-satellite information every second.Source: second window against TLE-based… · p. 4

identification result every second, whereas existing scheduling-cycle-based methods can identify

StarRTT integrates per-second serving-satellite identification with multi-step RTT prediction, while RT-LSTM combines historical and TLE-propagated future features. Under strict real-time back-testing, it achieves lower MAE than the three stated baselines.

identification result every second, whereas existing scheduling-cycle-based methods can identify

  • StarRTT integrates per-second serving-satellite identification and multi-step RTT prediction in one real-time pipeline, eliminating the identification blind zone.The framework enables continuous real-time RTT prediction throughout each scheduling cycle.Source: identification result every second,… · p. 5
  • RT-LSTM combines historical RTT and serving-satellite features with TLE-propagated future features through dual LSTM branches and gated fusion.It predicts RTT per second over the next 15 seconds; the identification algorithm achieves a minimum 98.5 % agreement with the full-cycle final identification result.Source: identification result every second,… · p. 5
  • Across Starlink V1.0, V1.5, and V2.0-mini, the analysis found distinct RTT characteristics within and outside an empirically identified bent-pipe coverage area, motivating gateway-relative features.The paper reports incorporating these features in RT-LSTM.Source: identification result every second,… · p. 5
  • 1.405 ms MAE averaged across all 15 prediction horizons and 15 reference time-slot indices; RT-LSTM reduces MAE by 4.3 %, 11.6 %, and 22.1 % versus LSTM, StarNet, and ASTM, respectively.The result is reported under strict real-time back-testing.Source: identification result every second,… · p. 5

II. PROPOSED REAL-TIME SERVING-SATELLITE IDENTIFICATION AND RTT PREDICTION

The proposed method updates serving-satellite identification every second and refines its trajectory scoring over the scheduling cycle. The section also describes the RT-LSTM framework and an assumption that bounds the identification method’s scope.

A. Proposed Serving-Satellite Identification Method

  • Per-second serving-satellite identification replaces three-point trajectory scoring with a refined scoring mechanism informed by temporal importance.The conventional method uses only start, midpoint, and endpoint observations; the proposed method updates identification every second.Source: A. Proposed Serving-Satellite… · p. 5
  • The algorithm assumes no handover after scheduling instant T until the next scheduling instant.It relies exclusively on angular distance from T through T+4.Source: A. Proposed Serving-Satellite… · p. 6
  • Dynamic beam switching upon serving-link obstruction can occur at arbitrary times, unlike regular scheduling-instant handovers.This is a distinct handover mode noted alongside the algorithm’s scheduling-cycle assumption.Source: A. Proposed Serving-Satellite… · p. 6
  • RT-LSTM uses dual LSTM branches and gated fusion to predict RTT over a 15-second horizon.The branches encode historical RTT, serving-satellite, and gateway-relative features, and TLE-propagated future features.Source: A. Proposed Serving-Satellite… · p. 7
  • The score combines angular distance and directional difference between observed and TLE-propagated satellite trajectories.Angular distance is used early, when directional estimates are distorted by limited obstruction-map resolution; directional difference is added as more trajectory points accumulate.Source: A. Proposed Serving-Satellite… · pp. 6–7

B. RT-LSTM Prediction Architecture

RT-LSTM combines historical and future LSTM representations through gated fusion to predict RTT over a 15-second horizon.

B. RT-LSTM Prediction Architecture

1) Real-time serving-satellite and gateway-relative feature encoding:

RT-LSTM forms a historical sequence from observations available through reference time t, then encodes it with a two-layer historical LSTM.

1) Real-time serving-satellite and gateway-relative feature encoding

  • At each historical time step from t−44 through t, the input combines RTT observations with serving-satellite and gateway-relative features.The features characterize geometric relationships among the serving satellite, user terminal, and gateways.Source: 1) Real-time serving-satellite and… · p. 8
  • The historical input window spans 45 seconds, matching the historical context duration used by ASTM.The passage states that ASTM’s context comprises 3,000 time steps at 0.015-second temporal resolution.Source: 1) Real-time serving-satellite and… · p. 8
  • A two-layer LSTM encodes the historical sequence, with its final hidden state representing the historical output.The final hidden state is the historical LSTM representation.Source: 1) Real-time serving-satellite and… · p. 8

2) Gateway-relative feature augmentation:

Gateway-relative features augment RT-LSTM’s inputs to represent RTT differences associated with bent-pipe and ISL link modes.

2) Gateway-relative feature augmentation

3) Future feature encoding and gated fusion:

RT-LSTM combines historical information with TLE-propagated future features of the identified serving satellite, then fuses the two representations to predict RTT for the next 15 seconds.

3) Future feature encoding and gated fusion:

  • The future sequence uses TLE-propagated features of the identified serving satellite from t+1 to one second before the next scheduling instant b(t).b(t) is the first scheduling instant after t.Source: 3) Future feature encoding and gated… · pp. 8–9
  • A future LSTM encodes this sequence, and the historical and future hidden states are projected into a common latent space.The projection uses learnable matrices and corresponding bias vectors.Source: 3) Future feature encoding and gated… · p. 9
  • A feature-wise gate controls the respective contributions of the historical and future representations before producing RTT predictions for the next 15 seconds.The gate weights the historical representation by g_t and the future representation by 1 − g_t.Source: 3) Future feature encoding and gated… · p. 9
  • The input features include RTT statistics, serving-satellite geometry and motion, and geometry relative to four bent-pipe gateways.RTT features include normalized RTT, its delta, and 3-second rolling mean and standard deviation; satellite features include position, angles, distance, relative velocity, time to exit the UT field of view, and days since launch.Source: 3) Future feature encoding and gated… · pp. 9–10

A. Experimental Environment

The experiments use continuous Starlink measurements collected in Seoul and compare identification results with trajectory-based baselines. The proposed real-time identification agrees with the full-cycle result at least 98.5% of the time.

A. Experimental Environment

  • The study collected approximately 114.5 hours of continuous measurements at Korea University in Seoul, South Korea.The passage also says timestamps were synchronized using Google’s NTP server every five minutes.Source: A. Experimental Environment · p. 10, ment across all datasets, timestamps… · p. 10
  • RTT was measured from the Starlink Router to the Tokyo-region AWS Point of Presence during a continuous, single-stream TCP downlink without bandwidth limitation.The work targets the next 15 seconds of RTT, updated every second, and evaluates models under strict real-time data-availability constraints.Source: ment across all datasets, timestamps… · p. 11

ment across all datasets, timestamps were synchronized with Google’s Network Time Protocol

  • The proposed identification method’s average three-point trajectory error score is 7.1°, versus 25.5° for StarNet and 12.7° for Ahangarpour et al..A lower score indicates a closer match between observed and TLE-propagated trajectories.Source: ment across all datasets, timestamps… · p. 11
  • At every second within a scheduling cycle, real-time serving-satellite identification agrees at least 98.5% with the result obtained from the full 15-second trajectory.This agreement is reported as evidence of reliable identification without waiting for the full trajectory.Source: ment across all datasets, timestamps… · p. 12

B. Empirical Analysis of Link Mode and RTT

Across satellite generations, measured RTT differs between bent-pipe coverage and inferred ISL cases. Under strict real-time back-testing, RT-LSTM has lower horizon-averaged MAE than the baselines at every reference time-slot index.

B. Empirical Analysis of Link Mode and RTT

  • V1.5 and V2.0-mini have inferred bent-pipe-to-ISL ratios of 21:1 and 4:1, respectively, while measured RTT averages about 35 ms in the bent-pipe coverage intersection and above 40 ms for inferred ISL cases.The analysis defines bent-pipe coverage using minimum gateway and UT elevation angles of 35° and 25°, respectively.Source: B. Empirical Analysis of Link Mode and… · p. 12

C. RTT Prediction Results

1) Advantage of real-time serving-satellite information:

  • 4.3%, 11.6%, and 22.1% are the MAE reductions for RT-LSTM versus LSTM, StarNet, and ASTM, respectively, averaged over all 15 horizons and 15 reference indices.RT-LSTM achieves lower horizon-averaged MAE than every baseline at every reference time-slot index.Source: 1) Advantage of real-time… · p. 13

2) Performance variation across reference time-slot indices and prediction steps:

The analysis reports RT-LSTM’s real-time benchmark result alongside changes in MAE near the next scheduling instant and correlations between gateway-relative geometry and RTT. It also describes multi-generation analysis and clarifies that bent-pipe and ISL modes are inferred for analysis, not treated as ground-truth routing.

2) Performance variation across reference time-slot indices and prediction steps:

  • MAE generally rises as the reference time advances toward the next scheduling instant, and increases sharply at t+15 for reference time-slot index 12, where future serving-satellite information is unavailable.A similar increase occurs at the prediction step crossing the next scheduling instant for other reference time-slot indices.Source: 2) Performance variation across… · pp. 13–14
  • Bent-pipe and ISL modes are inferred primarily from geometric constraints, with RTT distributions as supporting evidence; they are not ground-truth routing decisions.These inferred modes are used only for analysis.Source: 2) Performance variation across… · p. 13

3) Gateway-relative features:

  • Satellite azimuth relative to Otaru and distance to Yamaguchi rank first and second in correlation magnitude with RTT, at 0.264 and 0.257, respectively.The analysis describes gateway-relative geometry as closely associated with RTT variation across link modes.Source: 3) Gateway-relative features · p. 14

IV. CONCLUSION

  • Analysis across Starlink V1.0, V1.5, and V2.0-mini identified a bent-pipe coverage area and distinct RTT characteristics between bent-pipe and ISL modes.The conclusion describes StarRTT as addressing blind-zone limitations of scheduling-cycle-based serving-satellite identification.Source: §IV · p. 14

tions, we proposed RT-LSTM within this real-time framework, which jointly leverages historical

  • 1.405 ms MAE averaged across all 15 prediction horizons and 15 reference time-slot indices, with reductions of 4.3 %, 11.6 %, and 22.1 % versus LSTM, StarNet, and ASTM, respectively.RT-LSTM combines historical RTT, real-time serving-satellite and gateway-relative features, and TLE-propagated future features through historical and future LSTM branches with gated fusion.Source: tions, we proposed RT-LSTM within this… · p. 14
Loading 2610.32540v1…