Graph Neural Network Based Spatiotemporal Traffic Prediction Framework for Smart Communication Networks

Spatiotemporal traffic prediction has become a cornerstone of intelligent transportation systems within smart communication networks. Accurate forecasting of traffic flow enables congestion mitigation, real-time routing, and dynamic resource allocation. Conventional time-series or convolution-related models may not be able to capture the non-Euclidean nature of urban road networks and the intricate time-dependent interdependencies of road traffic. This paper presents an original Graph Neural Network (GNN) architecture, named ST-GWNet +, a combination of Graph Wavelet Embeddings, Temporal Transformers, Edge Memory Fusion and Dual-Level Attention, which can be used to capture complex spatiotemporal interactions. We achieve local traversals of fine-grained traffic transition and long-range temporal dependencies of our architecture by combining spectral and temporal learning units. When evaluated on a real urban traffic dataset, ST-GWNet++ has a Mean Absolute Error (MAE) of 3.22, RMSE of 4.38, and an R2 score of 0.918, surpassing the best baselines, namely, DCRNN, STGCN, and GMAN. The model proposed is also highly interpretable and has high generalization capabilities of time horizons. The piece of work prepares the way to more adaptive and explainable traffic management solutions and provides a scalable edge deployment of smart cities.