Transformer‐Based Large‐Scale and Intelligent Network Traffic Prediction and Optimization

Nowadays, this society relies more and more on large‐scale intelligent networking to operate the desired functions, which leads to a great and tremendous amount of Internet traffic, particularly at peak time. Such exponential traffic increase has caused a great challenge to network infrastructure, which in turn reflects the importance of network traffic prediction. Reliable traffic prediction can help the Internet service provider to manage their resource efficiently, so as to guarantee the service quality even at high demand and prevent the network congestion from happening. However, with the access to a tremendous smart devices, the corresponding traffic grows exponentially. In this way, it becomes vitally important to accurately capture and predict such traffic status. Traditional prediction models are usually applied to short‐term prediction scenarios, which are not suitable for real‐world scenarios, because the traffic is more complex. On the other hand, deep learning has been frequently used for network prediction due to its non‐linear modeling capability. Nevertheless, these methods may encounter trouble when dealing with the problems related to the long dependence relationship and dynamic space relevance among traffic data with a large amount of data. To address these challenges well, we propose a novel prediction architecture using the transformer structure. It takes the time and space factors into consideration when fulfilling traffic prediction. Specifically, on one hand, we separate the input sequence along the timeline, so as to better capture the dynamic space relevance to traffic. For each part of the captured sequence, we build the sub‐model for relevance modeling. Then, with these discrete traffic models with time and space features, we introduce the multi‐head attention mechanism to integrate them, so as to finally build the perfect relevance matching among the local and global traffic space. Our experiments indicate that the proposed transformer‐based architecture implement a highly accurate traffic prediction model while reducing the training time. Compared to the state‐of‐the‐art methods, the proposed one achieves high performance in terms of the mean absolute error, root mean square error, R‐squared, and efficiency in training time.