UL-CSI Data Driven Deep Learning for Predicting DL-CSI in Cellular FDD Systems

Frequency division duplex (FDD) systems dominate current cellular networks due to its advantages of low latency and strong anti-interference ability. However, the computation and the feedback overheads for predicting the downlink channel state information (DL-CSI) are the major bottlenecks to further improve the cellular FDD systems performance. To deal with these problems, in this paper, a convolutional long short-term memory network (ConvLSTM-net)-based deep learning method is proposed for predicting the DL-CSI from the uplink channel state information (UL-CSI) directly. In detail, our proposed ConvLSTM-net consists of two modules: one is the feature extraction module that learns spatial and temporal correlations between the DL-CSI and the UL-CSI, and the other one is the prediction module that maps the extracted features to the reconstructions of the DL-CSI. To evaluate the outperformance of the ConvLSTM-net, a long short-term memory network (LSTM-net) and a convolutional neural networks (CNN)-based schemes are simulated for comparisons. The simulation experiments consist of two parts. One part is that the hyper parameters of the proposed ConvLSTM-net are analyzed to explore their effects on the prediction performance. Another part is that experiments are conducted in the time domain and frequency domain, respectively, for selecting a more proper domain to predict the DL-CSI accurately. From the experiment results above, it can be verified that the proposed ConvLSTM-net with proper hyper parameters outperforms the compared schemes at predicting DL-CSI according to UL-CSI in the cellular FDD systems, especially in the time domain.