A Hybrid LSTM-Transformer Approach for Accurate Time Series Traffic Prediction in Cellular Networks

Accurate cellular traffic prediction is crucial for capacity management, energy efficiency, and network slicing in 5G networks. This study proposes a hybrid model based on Long Short-Term Memory (LSTM) and Transformer to predict network traffic using real operational Key Performance Indicators (KPI) data from 23 cells in Bandung over four months. This architecture combines the LSTM’s ability to capture local temporal dependencies with the self-attention mechanism to identify global patterns, enriched by sinusoidal time embedding and representation fusion through a concatenation layer. Through systematic hyperparameter optimization, the model achieves R2= 0.9357 and Root Mean Squared (RMSE) = 0.9822, outperforming baselines such as Autoregressive Integrated Moving Average (ARIMA), LSTM-only, and Convolutional Long Short-Term Memory (ConvLSTM). Ablation studies prove the significant contribution of each architecture component, while robustness analysis demonstrates stability under incomplete data conditions. The novelty of the research lies in the application of a hybrid architecture to real KPI data, rather than simulated grid datasets. These results provide a strong basis for the application of traffic prediction in operational scenarios, such as base station sleeping and dynamic resource allocation. For further research, the integration of external factors and the development of lightweight models for edge deployment are highly recommended.