Spatiotemporal network traffic forecasting using FFT-enhanced inputs and a ConvNeXt3D-mamba framework

Area of Science:

• Computer Science
• Artificial Intelligence
• Network Engineering

Background:

• Accurate grid-based network traffic forecasting is crucial for efficient network management and resource allocation.
• Existing methods often struggle with capturing complex spatio-temporal dependencies and spectral characteristics of network traffic.

Purpose of the Study:

• To develop and evaluate a novel deep learning framework for enhanced grid-based network traffic forecasting.
• To investigate the efficacy of integrating Fast Fourier Transform (FFT) features with ConvNeXt3D and Mamba architectures.

Main Methods:

• A learnable FFT-enhanced ConvNeXt3D-Mamba framework was proposed.
• A four-channel representation incorporating standardized residuals, FFT magnitude, and phase descriptors was constructed and refined.
• Spatial encoding was performed by ConvNeXt3D, and temporal modeling by Mamba for next-step prediction.

Main Results:

• The model achieved state-of-the-art performance on the Milan dataset with MSE of 0.0152, MAE of 0.0880, RMSE of 0.1232, and R² of 0.9848.
• Ablation studies confirmed that combining original and FFT-derived inputs yields superior performance.
• Robustness tests demonstrated stable performance under input perturbations.

Conclusions:

• The proposed FFT-enhanced ConvNeXt3D-Mamba framework significantly improves network traffic forecasting accuracy.
• The frequency-enhanced representation demonstrates transferability across diverse time-series tasks, including regression and classification.
• This research offers a promising direction for advanced network traffic analysis and prediction.