• BCAG-Net: a novel multi-branch deep learning model for Cellular traffic prediction. • Combines Butterworth and Chebyshev II filters for superior noise reduction. • Elman RNN, causal CNN, and attention GRU capture short- and long-term patterns. • Outputs fused via NAR network for robust multi-horizon forecasting. • Achieves higher accuracy and generalization than conventional models. Accurate real-time network traffic prediction is critical for efficient resource management and service quality. This work proposes BCAG-Net, a hybrid filtering–deep learning model that integrates Butterworth and Chebyshev Type II filters for effective noise suppression, followed by a multi-branch architecture combining Elman RNN, CNN, GRU, and attention mechanisms. Raw traffic data are denoised, decomposed, and transformed into sliding-window sequences. The Elman branch captures short-term dynamics, CNN extracts local patterns, and the GRU-attention branch models long-term dependencies. Outputs are fused via a Non-Linear Autoregressive (NAR) network for final prediction. Extensive experiments show BCAG-Net achieves a mean absolute error of 0.142, significantly outperforming baseline models across RMSE, MAE, MSE, and R².
Shadhar et al. (Sun,) studied this question.