Spectrum Efficiency Prediction for Real-World 5G Networks Based on Drive Testing Data
Zheng Xing, Haoyun Li, Wenjie Liu, Zixiang Ren, Junting Chen, Jie Xu, Cai Qin
Abstract
This paper studies the problem of predicting the spectrum efficiency (SE) for massive multiple-input multiple-output (MIMO) empowered 5G networks based on the reference signal received power (RSRP) collected from the drive test (DT). This problem is challenging because there is no precise model between the RSRP and the SE. The SE not only depends on the RSRP, which only captures the statistic of the channel, but also the beamforming strategy of the serving base station (BS) and the interference from the neighboring cells, which are not measured at the 5G client. This paper adopts a model-assisted data-driven approach to develop a machine learning model for the SE prediction. Specifically, a joint interference and SE prediction network is built, demonstrating prediction improvement over pure data-driven neural networks. In addition, a classification-assisted SE prediction network is constructed, which substantially reduces the prediction error at the low SE regime with marginally compromising the total prediction error. It is found that the model-assisted approach generally enhances the SE prediction accuracy by 2% approximately over a purely data-driven approach.