Research On Time-varying Channel Prediction Method Based On Deep Learning In 5G High-speed Mobile Scenarios

In recent years,with the progress of science and technology,people's travel needs are increasing.High Speed Railway(HSR)has outstanding advantages such as high efficiency,safety and environmental protection.It has increasingly become the most popular means of transportation and has been widely developed and deployed in China.The current domestic high-speed rail speed can reach 300km/h,and with the application of the 5G communication system,the high-speed rail speed will reach 500km/h or even higher.In the future 5G HSR environment,serious Doppler frequency shift will be caused due to the high-speed running of trains,which will cause rapid time change in the channel.In a fast time-varying channel,due to the presence of feedback and processing delays,the traditional channel state information obtained through channel estimation methods will be outdated,and can no longer accurately describe the current channel state,which will lead to the performance of the HSR communication system reduced seriously.Therefore,in order to ensure the communication quality in the 5G HSR environment,the channel prediction method is desired for channel information acquisition in this scenario.Considering that the 5G HSR scenarios will face higher moving speeds,the HSR wireless communication systems will face more severe Doppler shift at this time,and traditional channel prediction methods cannot achieve great performance in fast time-varying channel.To improve the accuracy of time-varying channel prediction in 5G HSR scenarios and reduce the complexity of channel prediction,this thesis explores an efficient time-varying channel prediction method which is more suitable for 5G HSR scenarios based on existing channel prediction methods.The main content and innovations are as follows:(1)For high-speed mobile orthogonal frequency division multiple access(OFDMA)systems,a new time-varying channel prediction method based on back propagation(BP)neural network is proposed.In order to avoid the influence caused by random initialization of network parameters,the proposed method first obtains a more ideal channel estimation based on data and pilot information,and uses it to pre-train the BP neural network to obtain the ideal initial network parameters;The initial value of the network obtained by pre-training is used to retrain the BP neural network with the channel estimation obtained based on the pilot to obtain the final channel prediction network model;finally,the single-time and multi-time prediction of time-varying channel is realized through online prediction based on the obtained prediction model.Theoretical analysis and simulation results show that the proposed method can significantly improve the accuracy of time-varying channel prediction and has lower computational complexity.(2)Aiming at the problems of falling into local optima and fixed network models in BP network training,a new type of time-varying channel prediction method based on deep representation learning via extreme learning machine(Dr-ELM)is proposed.Based on the single hidden layer neural network,in order to capture the deep information of the input data,the Dr-ELM is firstly used to extract the deep features of the channel from the historical channel,and the initial output weight of the network is obtained;then,in order to adapt to the change of the channel,the proposed method updates the output weights of the network in real time based on the newly constructed historical channel samples and the initial output weights,and obtains the current channel based on the updated output weights.Theoretical analysis and simulation results show that the proposed method has high prediction accuracy and is suitable for high-speed moving scenes.