Accurate prediction of signal strength and beam selection is essential for efficient 5G network operation. Although deep learning shows strong potential in both simulations and real-world network scenarios, achieving high accuracy requires large volumes of diverse, high-quality data. In this work, we propose a multi-task learning framework that jointly predicts Reference Signal Received Power (RSRP) and optimal beam index using real-world measurements from the Vienna 4G/5G Drive-Test Dataset. Unlike prior work focusing solely on how signals propagate, we extend the problem toward decisionmaking by incorporating beam prediction. We compare two architectures: a hard parameter-sharing model and a decoupled dual-branch model. Experimental results show that the proposed dual-branch model significantly outperforms the hardsharing approach, achieving an RMSE of 1.82 dB for RSRP prediction and improving beam selection accuracy to 75.2% (top-1) and 91.4% (top-2). Furthermore, we analyze the impact of spatial features by comparing configurations with raw GPS, no GPS, and noisy GPS inputs. The results show that adding GPS data only slightly improves performance. This suggests the model is learning how radio signals behave, rather than just memorizing specific locations.

Monika Velichkova; Antoni Ivanov; Philipp Svoboda; Agata Manolova; Vladimir Poulkov

2026 61st International Scientific Conference on Information, Communication and Energy Systems and Technologies (ICEST)

https://ieeexplore.ieee.org/document/11623577

5G New Radio, Beam Management, Deep Learning, Multi-task Learning, Radio Propagation Modeling