DuTSA: Dual-Branch Time Series Analysis Framework for Network Security Situation Forecasting

The increasing complexity of network threats necessitates proactive and adaptive forecasting solutions. However, existing network security situation forecasting (NSSF) methods often struggle to simultaneously capture long-term dependencies and short-term anomalies, resulting in suboptimal predictive performance. To address these challenges, we propose DuTSA, a dual-branch forecasting framework that combines a Transformer branch for global temporal modeling with a CBAM-enhanced convolutional branch for localized feature extraction. A hierarchical fusion mechanism integrates multi-scale features via concatenation and linear projection, while multi-layer stacking strengthens deep temporal representations. Moreover, a sliding-window strategy enables DuTSA to dynamically adapt to evolving threat patterns. Experimental results show that DuTSA consistently outperforms state-of-the-art models, achieving an 11% reduction in long-term forecasting error (MSE) and a 6.9% improvement in short-term prediction accuracy.