A Small, Fast, Quantized Transformer Based Neural Network for Bitrate Prediction
Sayandip Pal, Mahasweta Sarkar, Santosh V. Nagaraj · 2025
Bitrate prediction plays a crucial role in modern networks, particularly in satellite networks. Satellites dynamically adjust their throughput based on varying user demands and traffic patterns, making accurate bitrate forecasting essential for optimizing bandwidth and minimizing operational costs. This research aims to overcome the challenge by applying time series modeling techniques specifically for timeslot-based bitrate prediction. A novel GRU-LSTM ensemble model and a custom Transformer-based model are introduced to capture the intricate temporal dependencies present in network traffic. Emphasizing computational efficiency, the models are quantized to significantly reduce model size and inference time, making them suitable for real-time, resource-constrained applications. Evaluations show that the quantized GRU-LSTM model achieves a 3.55% MAPE and 0.66 ms inference time, outperforming individual GRU and LSTM models with higher MAPE values of 6.68% and 7.24%, respectively.