Predictive Channel State Information Feedback in Multiple Input Multiple Output Using Improved Recurrent Neural Network

In modern era, Multiple Input Multiple Output (MIMO) technology is a crucial element in wireless technology, where accurate prediction of Channel State Information (CSI) feedback is critical for beamforming, optimal resource allocation. However, due to traditional compression techniques based on quantization such as scalar and vector quantization which introduces errors in CSI feedback, which leads to inaccurate channel estimation. To overcome these challenges, an Improved Recurrent Neural Network (IRNN) based CSI feedback mechanism is proposed with an integrated self-attention mechanism to improve the prediction accuracy. Initially, a system based on downlink massive MIMO and Orthogonal Frequency Division Multiplexing (mMIMO-OFDM) is considered. This system consists of multiple gNodeBs (gNBs), which serve multiple User Equipments (UEs). These multiple gNBs transmits Reference Signals (RSs) to enable UEs, to estimate and provide CSI feedback. Furthermore, IRNN based channel predictors are deployed at both the gNBs and UEs, which learn the temporal relationships among CSI data and refine the feedback information. The proposed IRNN is improved by integrating self-attention mechanism. Finally, the proposed IRNN outperformed RNN by achieving better results in Normalized Mean Squared Error (NMSE) of -26.500, precoding gains of 0.998, cosine similarity of 0.998 and spectral efficiency of 1.425.