Physics-Guided Residual Learning for Radio Environment Map Reconstruction With Blind BS Estimation in Private 5G Networks

Abstract

Private 5G (P5G) deployments are increasing the risk of spectrum overlap and inter-network interference in enterprise and industrial environments. Accurate radio environment maps (REMs) are therefore essential for spectrum monitoring, coverage assessment, and regulatory compliance. In practice, however, REM reconstruction is challenging due to sparse, irregularly distributed drive-test (DT) reference signal received power (RSRP) measurements and incomplete or unreliable base station (BS) metadata. To address this problem, this paper proposes a physics-guided two-stage framework for REM reconstruction from sparse DT measurements. The first stage uses a farthestpoint diverse-elite particle swarm optimization (FP-DEPSO) algorithm for blind BS localization and propagation parameter estimation. The second stage uses a physics-guided randomforest-based residual framework (PG-RFR) to refine the baseline REM through site-specific residual learning from propagationaware and environment-aware features. In a challenging twosector scenario, the proposed framework achieves a threedimensional localization error of 9.84 m and a held-out REM reconstruction error of 5.210 dB, outperforming conventional interpolation-based and direct data-driven baselines.