To address the limitations of traditional path loss models,which fail to account for environmental information and perform poorly in cross-scenario and cross-band predictions,an environment-driven path loss prediction method for cross-band and cross-scenario applications is proposed.The method combines two-dimensional linear and rectangular environmen-tal features to describe the propagation environment and incorporates transfer learning into a random forest-based path loss prediction model.Two urban scenarios were constructed:Scenario 1 includes frequency bands of 140,220,280,and 300 GHz,while Scenario 2 focuses on the 140 GHz band.The method uses datasets at 140 and 220 GHz to predict path loss at 280 and 300 GHz and employs Scenario 1 data to predict Scenario 2 path loss.Results demonstrate that the proposed meth-od reduces the root mean square error(RMSE)for achieving cross-band predictions at 280 and 300 GHz by 3.331 1 and 4.321 5 dB and for cross-scenario predictions by 0.724 4 dB compared to methods without transfer learning optimization.
关键词
机器学习/亚太赫兹/路径损耗预测/随机森林/迁移学习
Key words
machine learning/sub-terahertz/path loss prediction/random forest/transfer learning
万方数据