Network traffic prediction model based on WT-TBiLSTM

Accurate short-term network traffic prediction is critical for proactive network security management, yet remains challenging due to the inherent complexity, noise, and non-stationarity of traffic data. To address this, we propose WT-TBiLSTM, a hybrid deep learning model integrating Wavelet Transform (WT) for noise reduction and a Topology-Attention-enhanced Bidirectional LSTM (TBiLSTM) for temporal feature extraction. First, the WT module decomposes raw traffic data using Daubechies8 wavelets, applies adaptive soft-thresholding to suppress noise, and reconstructs denoised signals through multi-scale analysis. The TBiLSTM module then captures bidirectional temporal dependencies while leveraging a topology attention mechanism to emphasize local structural patterns, enhancing sensitivity to abrupt traffic variations. Experiments on two real-world datasets (UK\_Academic and EU\_Core) demonstrate the model’s superiority. Compared to baseline models , WT-TBiLSTM reduces prediction errors by 64.37% RMSE, 65.45% MAE, and 67.89% MAPE on UK\_Academic, and by 37.26% RMSE, 39.35% MAE, and 38.76% MAPE on EU\_Co. The integration of wavelet denoising and topology attention significantly improves robustness against noise and captures multi-scale temporal dynamics, enabling precise short-term predictions. This work provides a practical solution for real-time network traffic forecasting, empowering administrators to preemptively mitigate security risks and optimize resource allocation.