With more 5G base stations coming into play, making an accurate assessment of RF-EMF exposure currently faces increasing demand to check if they meet regulatory requirements and ensure people’s safety. We present here PSF-Net, a novel deep learning network by uniting TabPFN’s meta-learned prior knowledge and SAINT’s dual attention structure; its use makes it particularly suitable to deal with applications like prediction of downlink power density and radiation level classification under different conditions within various kinds of 5G cell.
A major component in the design of this approach is an uncertainty-aware gating block that determines the optimal weighting for each model output—TabPFN or SAINT—based on the estimated prediction variance as quantified via Monte Carlo sampling during training or the prediction variance calculated using inference-time dropout. In addition, a residual multi-layer perceptron (MLP) is also included to extract refined fused features and maintain a steady gradient flow.
We evaluated the PSF-Net on a public data set of 3,624 georeferenced base stations, each of which has 34 features. Compared with Transformer, GNN, XGBoost and other strong baselines (Random Forest, Extra Trees, MLP, kNN, and SVR), the model shows improvement in all metrics – macro-averaged F1 and MAE and RMSE – for both classification and regression problems. Ablation studies have shown that the uncertainty gate and both encoder branches are important: removal of any one of them produces significantly worse performance. Further analysis of the calibration shows that the network tends to moderate overconfidence, especially on high exposure instances.
Taken as a whole, the study results suggest that PSF-Net is a practical means of achieving a reliable large-scale RF-EMF exposure assessment with both correct predictions and proper calibrated probabilities/uncertainty; moreover, it can provide useful inputs for how to advance safety-critical tabular modeling when facing difficult questions of safety on large datasets.