Lightweight AI-driven traffic forecasting and shaping for 6G LEO satellite networks

Abstract

Low Earth Orbit (LEO) satellite networks are expected to be a key enabler of 6G communications, providing global coverage and low-latency services for remote and underserved regions. However, their dynamic topologies, large-scale deployments, and limited onboard resources pose significant challenges to reliable service delivery. This paper presents a lightweight, AI-driven framework for service flow forecasting in LEO networks to minimize latency and ensure compliance with service level agreements (SLAs). Our main contributions are: (1) iTransformer_Lite, a resource-efficient transformer variant employing simplified embeddings, linear attention, and compact feedforward networks (CompactFFN) to reduce computational overhead; and (2) a multi-class Credit-Based Shaping (CBS) algorithm that leverages iTransformer_Lite predictions for dynamic, SLA-aware traffic shaping. Experiments on multiple public and satellite-specific datasets show that iTransformer_Lite achieves up to approximately 57 % lower memory footprint and approximately 2.3 × faster inference compared to the baseline iTransformer, while maintaining competitive or superior forecasting accuracy across diverse benchmarks.