Forecasting Achieves 94% Accuracy for LEO Satellite Beam Hopping Demand

Scientists are tackling the critical issue of predicting user traffic demand in rapidly expanding Low Earth Orbit (LEO) satellite networks. Yekta Demirci, Guillaume Mantelet, and Stephane Martel, from Polytechnique Montreal and MDA Space respectively, along with Jean-Francois Frigon and Gunes Karabulut Kurt et al., present a novel forecasting solution designed to mitigate the risk of buffer overflows and packet loss caused by unpredictable traffic bursts, a significant problem as these networks scale. Their research introduces key enhancements to existing forecasting models, improving accuracy by up to 94% and enabling more effective Beam Hopping (BH) technology for efficient resource allocation. This burst-aware approach isn’t limited to satellite communications, however, offering a broadly applicable solution for any wireless network grappling with intermittent, high-demand periods where precise forecasting is paramount.

This breakthrough addresses the potential for buffer overflows and packet loss that can occur when unexpected demand surges coincide with degraded link conditions in these networks. The study unveils a burst-aware transformer architecture, enhancing existing models to better anticipate and manage fluctuating user needs.

The core of this innovation lies in three key enhancements integrated into a transformer architecture. Firstly, the researchers incorporated a ‘distance from the last burst’ embedding, allowing the model to capture the proximity of upcoming bursts and improve prediction accuracy. Secondly, two additional linear layers were added to the decoder, enabling the forecasting of both upcoming bursts and their anticipated impact on network resources. Finally, an asymmetric cost function was implemented during model training, specifically designed to better capture the dynamics of burst events and refine the forecasting process.
These combined improvements allow for more precise anticipation of traffic fluctuations. Experiments conducted within an Earth-fixed cell under high-traffic conditions demonstrate the effectiveness of this approach. The proposed model not only reduces prediction error significantly but also maintains its ability to accurately capture bursts even at longer prediction horizons, as measured by the Mean Square Error (MSE) metric. This sustained accuracy is crucial for proactive resource allocation and preventing network congestion. The work establishes a new benchmark for traffic forecasting in LEO satellite networks, paving the way for more reliable and efficient communication services.

This research is particularly relevant given the increasing reliance on LEO constellations to bridge the digital divide and complement terrestrial networks. The ability to accurately forecast demand is essential for optimising spectrum and power management through beam hopping, dynamically allocating resources to areas with the highest need. Furthermore, the proposed solution is broadly applicable to any wireless network characterised by bursty traffic patterns, extending its potential impact beyond satellite communications. The. This breakthrough delivers improved performance in forecasting upcoming bursts and their impact on network resources, utilising a modified transformer model.
Experiments involved generating a high-demand traffic dataset comprising 60,000 samples, with burst points labelled using parameters k = 128 and h = 2.5. The team implemented modifications on the Informer model, a transformer architecture employing multi-head “Prob-sparse” self-attention, which reduces computational complexity to O(LQ log LQ) compared to the quadratic complexity of full attention. The model incorporates a distance from the last burst embedding, two additional linear layers in the decoder, and an asymmetric cost function during training to capture burst dynamics. Measurements confirm that the proposed model reduces prediction error by up to 94% at a one-step horizon, as quantified by the Mean Square Error (MSE) metric.

Data shows that the model maintains the ability to accurately capture bursts even near the end of longer prediction horizons, with MSE values ranging from 0.07, 0.17 for bursts occurring at a prediction length of 48. Table II details forecasting accuracy across prediction lengths of 1, 12, 24, and 48, demonstrating a 28% improvement in forecasting sequences containing at least one burst at a length of 48. Tests prove that the model outperforms benchmark models like Informer, ARIMA, and FARIMA, particularly in capturing burst traffic. An ablation study, detailed in Table III, quantifies the incremental impact of each enhancement, positional embedding, fully connected layers, and the asymmetric cost function, demonstrating that each modification improves burst forecasting ability. The proposed model demonstrably reduces prediction error, achieving up to a 94% reduction at a one-step horizon and maintaining accuracy even for bursts occurring near the end of longer prediction horizons, with Mean Square Error (MSE) values ranging from 0.07 to 0.17 for a 48-step prediction length. Ablation studies confirm that each enhancement, positional embedding, fully connected layers, and the asymmetric cost function, contributes to improved burst forecasting ability, and consequently, to better overall forecasting accuracy. The authors acknowledge that future work will focus on exploring lighter transformer variants to facilitate scalable deployment in resource-constrained environments, such as those found in LEO satellite networks.