Spatio temporal attention mechanism for real time cellular traffic prediction

Accurate prediction of cellular network traffic is a fundamental requirement for ensuring efficient network performance in the context of increasing mobile data demands. Existing models such as Convolutional Neural Network-Long Short Term Memory (CNN-LSTM), Temporal Fusion Transformer (TFT), and Reslearn often lack the capacity to capture the complex spatio and temporal patterns and variability inherent in real-world traffic, thereby limiting their effectiveness in practical deployments. This study proposes a lightweight hybrid framework incorporating a Spatio Temporal model with Attention Mechanism (STAM) to address these limitations and enhance predictive performance. The proposed model is trained on real-world cellular network data and is designed to capture both short-term fluctuations and long-term temporal dependencies. The attention mechanism embedded within the architecture allows the model to selectively focus on salient temporal features, improving its ability to learn meaningful traffic patterns while maintaining computational efficiency. Evaluation with standard metrics such as Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Symmetric Mean Absolute Percentage Error (SMAPE), and R 2 demonstrates improved prediction accuracy compared to traditional baseline models. The resulting predictions provide actionable insights for dynamic resource allocation and informed network planning. These capabilities support reduced latency, improved traffic distribution, and efficient bandwidth utilization, thereby contributing to enhanced Quality of Service (QoS), Spectrum Efficiency (SE), and Network Utility (NU) within next-generation cellular systems.