Digital Twin Empowered Channel Prediction for Massive MIMO Communications

Channel prediction techniques play an essential role in assisting the upcoming sixth-generation (6G) communication and smart applications, which can provide the path, speed, location, and channel conditions of the mobile devices by predicting the channel state information (CSI). Although the ray-tracing (RT) technique can help construct deterministic channel models with precise predictions, it cannot fully recover the real radio propagation scenes due to the massive parameters involved, and may lead to prediction inaccuracy. To this end, we propose a digital twin-empowered channel prediction framework for massive multiple-input multiple-output communications. Both the measured data and RT generated data are fed to the conditional generative adversarial network (CGAN) for model training, where the RT data acts as the conditioning information to assist data generation and discrimination to improve the accuracy of channel prediction. The Transformer model is integrated into CGAN to better handle the time series data. Experiments show that the proposed scheme can improve the prediction performance, which increases the channel correlation and reduces the normalized mean square error by at most 13% and 6 dB, respectively, as compared with popular existing schemes.