PREBAS: Preemptive Bandwidth Scaling for WebRTC in LEO Satellite Networks Using Lightweight Neural Prediction

Low Earth Orbit (LEO) mega-constellations promise universal broadband, but their orbital mechanics — terminals moving at 7.6 km/s, with handoffs recurring every few minutes — produce abrupt capacity crashes that violate the stationary-channel assumptions behind modern real-time protocols. Google Congestion Control (GCC), the standard for Web Real-Time Communication (WebRTC), only detects congestion after a queue has already built, so during a LEO handoff the encoder keeps pushing a high bitrate into a channel that has already collapsed, producing severe bufferbloat and, in a typical five-minute session, more than 1,000 frame freezes. This paper introduces PREBAS (PREemptive BAndwidth Scaling), an edge-native predictive framework built around an ultra-lightweight one-dimensional convolutional neural network (1D-CNN). Reading a 320 ms sliding window of capacity, round-trip time (RTT), and loss-burst telemetry, the model isolates the causal “RTT creep” signature that appears 200–500 ms before a handoff, then preemptively scales bitrate down by a factor of Sf = 0.40 to drain the transmission queue before the crash, with a 300 ms cooldown preventing GCC from immediately reversing the cut. Across 30 independent stochastic trials — covering handoffs, ITU-R P.618-14 rain fade, and channel noise — PREBAS reduces frame freezes by 79% relative to GCC at a cost of only 8.9% of average bitrate, lifting the ITU-T P.910 QoE score by 8.36 points (43% relative) to within 2.9% of the Oracle ceiling. An ablation confirms the result depends on reading capacity, RTT, and loss jointly, filtering 92% of rain-fade false positives a single-signal heuristic misses. At 1,537 parameters and 0.091 ms inference latency on commodity ARM hardware, the model is deployable on existing edge infrastructure without GPU acceleration.

• Status: accepted for the 2026 IEEE RTC in Chicago. The acceptance rests on the organiser's notification, which is available on request; the submission portal row is not a public URL, so it cannot be checked from this page.
• Accepted is the claim this page supports — not published. The RTC research track is published in IEEE Xplore, but this paper appears in neither IEEE Xplore nor Crossref as of 28 September 2026. The conference was held 22–23 September 2026 and proceedings normally follow by weeks or months, and inclusion depended on author registration, which was not completed: the associated mailbox holds only registration reminders, with no registration or payment confirmation. Treat any description of this paper as already in IEEE Xplore as premature.
• No artifact was released. The paper reports measurements but its code, data and raw outputs are not public, so nothing on it can be independently reproduced — which is what the Artifact label below means, not a judgement about the work.

No artifact released. This paper reports measurements, but its code, corpus and raw outputs are not public, so the numbers above cannot be independently checked. Treat the results as unreproduced.

@misc{poudel2026prebas,
  title  = {PREBAS: Preemptive Bandwidth Scaling for WebRTC in LEO Satellite Networks Using Lightweight Neural Prediction},
  author = {Sushant Poudel},
  year   = {2026},
  note   = {2026 IEEE RTC — Chicago, USA},
  url    = {https://sushantpoudel2028.com.np/papers/prebas.pdf}
}

Claims about his own work that the ledger covers are re-checked weekly against their public source by sushant-me/reputation — 27 claims today, 19 checked live. Coverage is stated rather than implied: it does not yet extend to every manuscript listed here. The reproducibility note above is part of that record: where no artifact exists, the page says so instead of implying one.