Spatially validated physics-guided residual learning for indoor path loss prediction

Abstract

Accurate indoor path loss prediction requires models that remain physically interpretable while capturing the structured variability introduced by walls, clutter, and non-line-of-sight propagation. This paper presents a physics-guided residual-learning framework for indoor path loss prediction in which a parsimonious per-band large-scale baseline is first calibrated under a leakage-aware nested spatial validation protocol, and machine learning is then used to model the remaining structured error through geometry-aware descriptors derived from a 2D floor-plan abstraction. Measurements were collected in a complex real indoor environment at five frequency bands between 0.7 and 7.0 GHz under LOS and NLOS conditions. The LS baseline provided a robust and interpretable first approximation, while residual learning consistently improved prediction on an untouched outer spatial test region. Regularized linear residual learners yielded modest gains, indicating that the remaining error contains only a weak linear component. In contrast, nonlinear ensemble models achieved substantially stronger improvements, with Random Forest providing the best validation-driven performance and XGBoost ranking a close second. SHAP analysis showed that the residual correction is driven mainly by local wall structure around the propagation path, receiver-local clutter, path-proximal geometry, and diffraction-inspired angular descriptors, rather than by frequency alone. The results support the view that a spatially validated hybrid strategy can improve indoor path loss prediction while preserving interpretability and reducing the risk of optimistic bias due to spatial leakage.