Joint Traffic Prediction and Large Flow Detection Using a Lightweight LSTM-Attention Hybrid Model

Abstract

With the rapid development of cloud computing and network communication, real-time monitoring and accurate detection of large flow events in server traffic have become crucial for ensuring stable network operation and efficient resource scheduling. However, existing traffic detection methods often suffer from poor ability to capture long-term temporal dependencies, low discrimination between large flows and transient bursts, and high computational overhead, making them difficult to adapt to resource-constrained deployment scenarios. To address these issues, this paper proposes a lightweight LSTM-Attention hybrid model for joint traffic prediction and large flow detection. The model adopts a single-layer LSTM encoder to efficiently extract temporal features from 6-dimensional multivariate traffic sequences, and introduces a position-biased single-head attention mechanism to enhance the perception of long-duration large-flow patterns. A multi-task output layer is designed to simultaneously realize traffic regression prediction and three-class classification (normal flow, burst flow, large flow), with a weighted hybrid loss function balancing the two tasks. Comprehensive experiments on real-world server traffic datasets demonstrate that the proposed model outperforms state-of-the-art baselines in terms of prediction accuracy, detection F1-score, and inference efficiency. Ablation studies verify the effectiveness of each core component, confirming that the model can achieve high-precision and real-time large flow detection with low computational cost, providing a reliable solution for practical network management.