A self-attention sequential model for long-term prediction of video streams

Video traffic prediction is a key technology to achieve accurate transmission bandwidth allocation and improve the quality of the Internet service. However, the inherent high rate variability, long-term dependence and short-term dependence of video traffic make it difficult to make a quick, accurate and long- term prediction:because existing models for predicting sequence dependencies have a high complexity and prediction models fail quickly. Aiming at the problem of long-term prediction of video streams, a sequential self-attention model with frame structure feature embedding is proposed. The sequential self-attention model has a strong modeling ability for the nonlinear relationship of discrete data. Based on the difference of correlation between video frames, this paper applies the time series self-attention model to the long-term prediction of video traffic for the first time. The existing time series self-attention model cannot effectively represent the category features of video frames. By introducing an embedding layer based on the frame structure, the frame structure information is effectively embedded into the time series to improve the accuracy of the model. The results show that, compared with the existing long short-term memory network model and convolutional neural network model, the proposed sequential self-attention model based on frame structure feature embedding has a fast inference speed, and that the prediction accuracy is reduced by at least 32% in the mean absolute error. © 2024 ournal of Xidian University. All Rights Reserved.