Uncertainty-Aware RSRP Prediction on MDT Measurements Through Bayesian Learning

Accurate and efficient propagation modeling is a key requirement for radio planning in cellular networks. Here, deep learning has recently shown promising performance in real-world evaluations but also requires an extensive amount of diverse training data to generalize well to unseen scenarios. In this work, we study the potential of using crowdsourced RSRP measurements from real-world MDT data to train deep learning-based propagation models. We utilize an uncertainty- aware Bayesian learning approach to adequately address the noisy characteristics of such data sources. This allows us to assess not only the achievable prediction performance, but also the uncertainty estimates which enable selective prediction. Our results show that - depending on the level of detail of the provided environmental data - a MAE of ≈ 6 dB can be achieved, with a further reduction to below 5 dB when removing samples with high aleatoric uncertainty. Meanwhile, the epistemic uncertainty reliably highlights scenarios not sufficiently captured in the training data, hence compensating for the collection bias.