A Hybrid CNN-LSTM Architecture for High Accurate Edge-Assisted Bandwidth Prediction

Abstract

In this letter, we propose an edge-assisted bandwidth prediction scheme based on a hybird convolutional neural network (CNN) and long short-term memory (LSTM) architecture. To accurately predict available downlink bandwidth with limited and regulated monitoring information, we decouple feature extraction and sequence learning to capture the feature vector by CNN and resolve the historical feature collection issue by LSTM, respectively. A two-stage training strategy is further adopted to improve the prediction accuracy and generalization ability. In addition, we deploy and evaluate the proposed bandwidth prediction scheme in our prototype system. The test results show that our proposed scheme improves the mean absolute error (MAE) and root mean square error (RMSE) by 81.4% and 82.0%, respectively, compared with the LSTM only prediction scheme, and consumes less than 3 ms running time, which is sufficient for many real-time video transmission applications.