Time Series Forecasting for Network Traffic

With the rapid development of intelligent networking technologies, time series forecasting models have become vital for network traffic management, enabling performance optimization and resource scheduling in large-scale network systems. This study proposes an adaptive multi-scale sequence kernel fusion model tailored for network traffic prediction tasks. By leveraging interface-level time series data from network devices, the model captures complex temporal patterns and cross-metric dependencies, addressing the limitations of traditional models that treat metrics independently. The model consists of three core modules: a Multi-scale Dynamic Modeling module, an Aggregation and Redistribution module, and a Multi-Expert Fusion module. Together, they enhance the model’s ability to represent periodicity, non-stationarity, and structural dependencies in heterogeneous traffic data. Experimental results demonstrate improved accuracy and robustness across diverse network environments.