| Increasing wireless data traffic and communication demands for new types of services bring new challenges to future communication systems in terms of key performance indicators such as data transmission rate,connection density,latency,and mobility.To meet more comprehensive application scenarios and higher performance requirements,future communication systems will explore higher frequency bands,be equipped with larger antennas,and require shorter delay,which bring difficulties in acquiring channel state information from different aspects such as accuracy,delay,complexity,guide transmission and feedback overhead.Therefore,the channel information of some of the channels can be used to infer other channel information,i.e.,channel prediction,to meet the demand for channel information in communication systems.The theoretical basis of channel prediction lies in the fact that channels experience the same propagation environment,but the complex mapping relationships between channels and environments often cannot be represented mathematically,and at the same time,the powerful approximation capability of deep learning can make continuous approximation to complex and nonlinear channel mapping relationships.Therefore,this paper explores a deep learningbased channel prediction method,analyzes the principles and shortcomings of existing methods for channel prediction,designs a neural network structure by combining the characteristics of MIMO channel data,and validates the method in terms of accuracy and generalization.The main contents of this paper are as follows:(1)Study of multi-channel parallel MIMO channel prediction methods.The advantages of deep learning in the uplink-to-downlink channel prediction problem and the problems in uplink-to-downlink channel prediction in MIMO systems based on deep learning are analyzed.Based on CNN,a channel prediction method from uplink channel to downlink channel is proposed by combining the spatial correlation between channel data.Based on the global correlation of channel data in the time,frequency,and antenna domains,the structure of a multi-way parallel channel prediction neural network is designed by combining the CNN structure with full convolutional layers.The channel data sets are generated for different scenarios,different user speeds,different antenna sizes,etc.The comprehensive performance of the uplink to downlink channel prediction method is verified by neural network error analysis and simulation of transmission.(2)Study of Massive MIMO system time-frequency-space joint channel extrapolation method.The advantages of deep learning in channel extrapolation problem are analyzed,and the problems of deep learning based channel extrapolation in each application scenario and the common problems in all application scenarios according to the principles of channel extrapolation in different application scenarios.Based on the global correlation of channel data in the time-frequency-space domain,a CNN-based joint channel extrapolation method in the time-frequency-space domain is proposed.Considering the performance of the multiple parallel uplink to downlink channel prediction method in multi-domain joint prediction,the neural network structure for the multi-domain joint channel extrapolation method is designed based on its neural network structure.The accuracy and stability of the multi-domain channel extrapolation method are verified by neural network error analysis and beam selection simulation.(3)Study of Massive MIMO upstream and downstream prediction and channel extrapolation joint prediction method.The advantages of the joint prediction method of upstream and downstream channel prediction and channel extrapolation are discussed and different joint prediction methods are proposed for the application scenarios with both uplink to downlink channel prediction and time-frequency air domain channel extrapolation requirements.After that,the different joint prediction methods are evaluated from the perspectives of prediction accuracy and application scenario scope through simulation comparison experiments,and the methods with obvious advantages are focused on,and their respective advantages and disadvantages as well as suitable usage scenarios are analyzed. |
