E-LDAC: A High-Accuracy Method for Predicting Bursty Latency in 5G Networks

The low latency of 5G networks greatly facilitates people's daily lives. However, the increases in latency caused by various factors, such as network configuration issues and environmental interference, may severely impact latency-sensitive applications. Therefore, accurately predicting network latency, especially bursty high latency, provides a new perspective for network operations and maintenance (O&M). Operators can proactively perform resource allocation and routing adjustments to prevent network faults. Nevertheless, network latency is a complex stochastic process, and the nonlinear correlations between network parameters and latency make latency prediction challenging. Imbalanced in latency further increases the difficulty of predicting bursty high latency. To address these challenges, we propose a latency prediction method called E-LDAC based on an LSTM model. First, by employing kernel density estimation (KDE), we improve the accuracy of the traditional training compensation process based on the empirical label density. Next, in the multiscale feature extraction stage, we realize the calibration of data features and improve the representation ability of data through multirate sampling and inverse transform-based interpolation. Then, to enhance the adaptability of the LSTM model to imbalanced data and explore the dynamic trends of latency, we design a jointly enhanced loss function with weight sharing. Finally, in a practical 5G multiservice network environment, we experimentally validate that the E-LDAC method achieves higher prediction accuracy than other methods in terms of overall latency, bursty high latency, and extended network fault prediction.