Path Loss Prediction Based on Dielectric-Aware Residual Learning

Abstract

Upper mid-band communication is a promising technology for future wide-area 6G networks. In this band, accurate path loss (PL) prediction is essential for coverage evaluation and network deployment. Existing AI-based PL prediction methods mainly use geometric environmental representations, but usually ignore the electromagnetic properties of building materials. In this paper, we propose a dielectric-aware residual learning framework for upper mid-band urban PL prediction. The proposed method constructs a dielectric constant map from the vector map to represent material-related electromagnetic properties and designs a dual-branch network to extract geometric and dielectric features separately. An empirical-modelassisted residual learning strategy is further adopted to correct the empirical-model prediction. Field measurement results at 7050 MHz show that the proposed method achieves an RMSE of 4.22 dB, outperforming the compared methods. Ablation and convergence analyses verify the effectiveness of the proposed scheme.