Cellular Network Traffic Prediction Based on Graph Enhanced Transformer in Dense Urban Scenarios

With the continuous increase in cellular network data, accurately predicting cellular network traffic has become crucial for network optimization and resource allocation. However, this task is highly challenging due to the complex spatiotemporal dependencies exhibited by cellular network traffic. To address these challenges, this paper proposes a model called the Graph Enhanced Traffic Transformer (GET-Transformer). Two core modules are designed for this model. First, we designed the Graph Enhanced Traffic Generator (GET-Generator), which uses a GCN as a preliminary extractor to initially capture spatiotemporal features across multiple cells, generating an initial cellular network traffic prediction map. Second, the features preliminarily extracted by the GET-Generator are embedded into a unified dimension to prepare for the subsequent multihead attention mechanism calculations. Third, we propose the Graph Enhanced Traffic Refiner (GET-Refiner), which further refines the initial prediction by leveraging the attention mechanism to accurately capture spatial dependencies between cells, thereby achieving precise predictions. Finally, our experimental results indicate that the GET-Transformer demonstrates up to a 58.56% performance improvement over traditional Transformers in terms of the MAE metric and up to a 52.47% improvement in the RMSE metric when applied to real cellular traffic datasets. Compared to GCN, the GET-Transformer also shows performance improvements of 46.58% and 58.79% in MAE and RMSE metrics, respectively.