چكيده به لاتين
With the rapid advancement of communication technologies and the increasing applications of cellular networks, especially with the emergence of fifth-generation networks, the volume and complexity of cellular network traffic have significantly increased. Efficient management of these networks requires accurate traffic prediction in both temporal and spatial dimensions, aiming not only to improve service quality but also to maximize network resource efficiency. This thesis introduces an advanced model based on deep learning and transformer networks for analyzing and predicting traffic in cellular networks. The proposed model consists of two main components: the global spatiotemporal block and the local spatiotemporal block. The global spatiotemporal block utilizes a hybrid structure that combines spatial and temporal transformers with graph convolutional networks to capture complex spatial correlations and long-term temporal dependencies, thereby enhancing prediction performance. The experimental results presented in this thesis, conducted using real-world traffic datasets from Telecom Italia, demonstrate that the proposed model significantly outperforms both traditional and modern methods in traffic prediction. When compared with methods such as GLSTTN, MVSTGN, ST-DenseNet, and LSTM, the proposed model exhibits superior performance in evaluation metrics, including Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). Specifically, the proposed model achieves a %3.45 reduction in error compared to the state-of-the-art model (GLSTTN). This improvement highlights the model's capability in extracting and modeling the complex, nonlinear features of cellular traffic. These advancements not only reduce errors but also enhance prediction accuracy across various cellular network conditions, positioning it as an innovative and efficient solution for cellular network traffic prediction.
