Enhancing Spatiotemporal Networks with xLSTM: A Scalar LSTM Approach for 5G Traffic Forecasting

Abstract

Accurate spatiotemporal traffic forecasting is vital for optimizing 5G networks. Traditional LSTM models struggle with capturing complex spatiotemporal dependencies, limiting predictive performance. To address this, we propose an enhanced Spatiotemporal Network (STN) integrating Scalar LSTM (sLSTM), a more efficient variant designed to improve temporal modeling while reducing computational complexity. Our dualpath STN processes the input through an sLSTM for sequential feature extraction and a three-layer Conv3D path for spatial feature learning, with both outputs fused in a dedicated fusion layer for enhanced spatiotemporal representation. By incorporating sLSTM, our model stabilizes gradients, accelerates convergence, and enhances accuracy. Experiments on real-world mobile traffic datasets show a 23% MAE reduction over ConvLSTM, with a 30% improvement on unseen data, demonstrating superior generalization for 5G traffic prediction.