Near-field Extremely large-scale MIMO Data Rate Prediction Based on Deep Learning

The channel matrix dimension of the near-field ultra-large-scale MIMO (Multiple-Input Multiple-Output) system is extremely high due to the significant increase in the number of antennas, thus aggravating the computational and storage burdens, and posing challenges to real-time processing and system resource management. Furthermore, traditional methods for obtaining Channel State Information (CSI) may perform poorly in near-field extremely large-scale MIMO systems, making it difficult to accurately capture the channel characteristics, which in turn affect the overall performance of the system. This study utilized the CsiNet-LSTM (Long Short-Term Memory) model to realize the channel capacity prediction.. This method combined the efficient CSI compression technique of CsiNet model with the temporal prediction capability of LSTM network, which could more accurately capture the dynamic characteristics of near-field extremely large-scale MIMO channels, thereby improving the accuracy of channel capacity prediction. During the research process, this article utilized communication simulation tools to generate CSI data under multiple propagation environments and normalize and segment them, then built encoders and decoders for the CsiNet model for extracting and reconstructing CSI features, and finally combined them with the LSTM model for time series modeling. The experimental results showed that the signal strength of the normalized signal strength of CsiNet-LSTM in a multipath propagation environment reaches 0.6, and the signal quality under noise conditions reached 0.7, which was superior to other models and demonstrated stability in complex environments. In terms of real-time performance, CsiNet-LSTM had an average prediction time of 0.35 seconds and a processing speed of 2857 samples per second, demonstrating excellent real-time processing capabilities compared to other models.