Research Article
A Reproducible Public-Trace Benchmark for Traffic Forecasting and Scheduling in Intermittent Space–Air–Ground Networks: Methodology and Potential Relevance to Smart-Grid IoT
@INPROCEEDINGS{10.1007/978-3-032-43712-9_6, author={Jilin Zhang and Zhenjiang Zhang}, title={A Reproducible Public-Trace Benchmark for Traffic Forecasting and Scheduling in Intermittent Space--Air--Ground Networks: Methodology and Potential Relevance to Smart-Grid IoT}, proceedings={Smart Grid and Internet of Things. 10th EAI International Conference, SGIoT 2026, Taichung, Taiwan, November 21--22, 2026, Proceedings}, proceedings_a={SGIOT}, year={2026}, month={11}, keywords={Space--Air--Ground Integrated Network Traffic Forecasting Task Offloading Intermittent Links Edge Computing Reproducible Benchmark Smart-Grid IoT}, doi={10.1007/978-3-032-43712-9_6} }- Jilin Zhang
Zhenjiang Zhang
Year: 2026
A Reproducible Public-Trace Benchmark for Traffic Forecasting and Scheduling in Intermittent Space–Air–Ground Networks: Methodology and Potential Relevance to Smart-Grid IoT
SGIOT
Springer
DOI: 10.1007/978-3-032-43712-9_6
Abstract
Intermittent satellite contacts and disruptions to aerial relays complicate cloud–edge–device task offloading in space–air–ground integrated networks (SAGINs). We present a reproducible public-trace benchmark that independently replays measured GÉANT traffic, CelesTrak Iridium-NEXT orbital elements, and NASA POWER weather observations in a three-tier service model while separating forecasting effects from allocation rules and information timing. LP-GRU obtains the lowest mean absolute error (MAE) of kbps, 7.05% below seasonal autoregression. Its 1.30% MAE reduction relative to SharedGRU is not conclusive over five initializations (paired mean difference kbps, 95% confidence interval (CI) kbps; ). At 20% outage severity, forecast-weighted LP-GRU completes 63.134% of task units, compared with 60.081% for binary round robin. However, no-forecast current-queue weighting reaches 63.159%, and pre-arrival backlog-only weighting reaches 63.195%. LP-GRU exceeds persistence by only 0.0063 percentage points (95% CI [0.0045, 0.0082]) in this setting. Across nine frozen-checkpoint deadline–capacity settings, its improvement over persistence is 0.000–0.028 points, whereas current-queue weighting exceeds LP-GRU by 0.001–0.157 points. Thus, when timely queue information is available, magnitude-aware allocation rather than forecasting explains most of the completion gain. The benchmark provides a controlled basis for studying predictive reservation under longer horizons, delayed observations, heterogeneous priorities, and application-specific IoT workloads.