Robust Long-Term Spatial–Temporal Forecasting for Dynamic Networks: The MDETST Model

Abstract

Accurate network traffic forecasting is essential for the proactive management of modern communication systems, particularly for emerging satellite constellations. However, forecasting network traffic is a non-trivial task, as the data exhibit complex statistical properties such as self-similarity, spatial–temporal long-range dependence, and non-stationarity, which challenge even the recent surge of deep learning models like iTransformer and DLinear. To address these limitations, this paper proposes MDETST, a novel Transformer-based architecture designed to more effectively capture and disentangle these complex, multi-scale spatio-temporal dependencies. Recognizing the scarcity of public satellite data, we conduct a comprehensive evaluation on the large-scale CESNET-TIMESERIES24 dataset, a challenging real-world benchmark that shares the critical statistical properties of both satellite and terrestrial systems. Extensive experiments demonstrate that our proposed MDETST model achieves state-of-the-art performance, consistently and significantly outperforming strong baselines, especially in demanding long-term forecasting scenarios. This work provides a robust and effective solution for complex traffic forecasting, showing strong potential for application in both satellite and terrestrial network management.