AC-LSTM: Adaptive clockwork LSTM for network traffic prediction

作者：Ge Yunfeng, Zonghuan Guo, Peng Wang, Xiaogang Li, Hongyan Li · 发表于：Journal of Information and Intelligence · 年份：2025 · DOI：10.1016/j.jiixd.2025.06.004 · 被引用次数：3 · 研究领域：Traffic Prediction and Management Techniques、Transportation Planning and Optimization、Music and Audio Processing

Accurate traffic load prediction on network links is essential for operators to optimize routing strategies based on predictive data. Deep time series analysis offers promising approaches to modeling network traffic flows. The traditional Long-Short-Term Memory (LSTM) model is commonly used to capture complex nonlinear relationships in temporal data. However, LSTMs often face challenges with long-range forecasting, either leading to diminished accuracy due to vanishing gradients or incurring high computational costs. To address these issues, this paper introduces the Adaptive Clockwork LSTM, a modified LSTM-based model that aims to provide accurate long-term predictions while reducing computational complexity. Dividing the LSTM unit into multiple segments that operate at adaptively varying intervals, an adaptive correlation mechanism is used for efficient sampling. The experimental results show that it outperforms existing methods, achieving a 20% improvement in forecast accuracy and a 65% reduction in computational time. These improvements make it a promising solution to reduce the costs associated with network management.