Abstract: The incomplete sample space of ray-tracing-data may increase high-prediction-error users in the massive multiple-input multiple-output channel amplitude prediction. To characterize the channel propagation features of all users, a method for 3D ray reconstruction is proposed based on extended probability distribution conditional variational auto-encoder (CVAE). The prior probability distribution is selected based on the sparsity of user ray samples. A new training set of ray samples is generated for high-prediction-error users by enhancing CVAE to make the latent variable distribution of ray-tracing-data fit the features of high-prediction-error users better. The simulation results show that the number of high-prediction-error users can be reduced to 53.59% by new training set based on the proposed method. Moreover, the new set improves the channel amplitude prediction accuracy by 7.8% while significantly reducing the time overhead of predicting the channel amplitude.

Key words: massive multiple-input multiple-output, 3D channel model, conditional variational auto-encoder, ray-tracing