Evaluation of Propagation Models and Machine Learning Through Path Loss Analysis in Different Environments Using Lora Technology at 915 MHz in the Department of La Paz, Bolivia

Abstract

The performance of Low-Power, Wide-Area Networks (LPWANs) is critical for long-range communication systems, but signal propagation in complex topographies limits their reliability. Theoretical propagation models often provide inaccurate path loss predictions in such environments, leading to suboptimal network design. This work presents an empirical evaluation of LoRa technology at <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 1 5 ~ M H z}$</tex> in La Paz, Bolivia, and compares the accuracy of classical propagation models against data-driven machine learning (ML) techniques. A dataset was constructed from a field measurement campaign conducted in urban (NLOS), suburban (mixed LOS/NLOS), and rural (LOS) scenarios. The Okumura-Hata, 3 GPP UMi, and COST231-WI models were evaluated. Among these, the 3GPP UMi model demonstrated the lowest error, yet yielded a Root Mean Square Error (RMSE) of 11.58 dB against measured data. In contrast, ML models trained on the same data showed higher accuracy. A Random Forest model achieved an RMSE of 3.49 dB, and an Artificial Neural Network model achieved an RMSE of 4.57 dB. The results indicate that ML models provide more accurate path loss predictions than theoretical models in complex terrains, offering an effective tool for planning LoRa-based communication networks.