This paper studies grid-based network traffic forecasting with a learnable FFT-enhanced ConvNeXt3D-Mamba framework. For the Milan dataset, we construct a four-channel representation from standardized residuals, FFT magnitude, and sine/cosine phase descriptors, and refine the spectral channels using descriptor projection and frequency-wise gating. The refined representation is spatially encoded by ConvNeXt3D along a preserved length- T axis and temporally modeled by Mamba for next-step prediction. On Milan, the proposed model achieves an MSE of 0. 0152, an MAE of 0. 0880, an RMSE of 0. 1232, and an \ (R²\) of 0. 9848 on the standardized residual scale, outperforming baseline models, structural variants, and recent comparison models under a unified protocol. Ablation results indicate that the combined original and FFT-derived input provides stronger overall performance than either source alone. Robustness tests under multiple input perturbations show stable performance. Auxiliary experiments on Madrid traffic, Electricity Load Diagrams 2011–2014, LargeST, and Wafer further support the transferability of the frequency-enhanced representation across regression and classification time-series tasks.
Zhang et al. (Thu,) studied this question.