An Effective Neural Network for Communication Signal Prediction via Propagation Path Serialized Representation

Communication signal prediction holds great significance for the optimization and deployment of B5G networks. In this letter, we propose a neural network model with Propagation Path Serialized Representation (PPSR) to predict the signal. Previous works treat the propagation path as a whole, ignoring the position sensitivity of propagation. For example, obstacles located in different positions should have varying effects. To solve this issue, PPSR proposes a serialized representation of the path. It divides the path into multiple propagation units. Each unit is represented by its geographical and position information, transforming the path into a position-sensitive sequence. PPSR then employs an effective sequence processing network model: Transformer, to derive the overall path representation from the propagation units. The Transformer can capture the correlation among different units along the path. Experiment results show that the proposed PPSR outperforms the existing methods in both prediction accuracy and model generalization ability, which mainly owes to the full consideration of the positional sensitivity and correlation among the propagation units on propagation paths.