SINR Prediction Methods Based On Machine Learning

Optimizing wireless communication system to ensure the stable and reliable transmission service quality is a significant work for practical applications.Downlink adaptive modulation and coding technology is one of the key technologies of wireless communication,which depends on the accurate and timely channel quality feedback of the receiver.However,there is a delay in the quality feedback process of the downlink channel in the real communication system,which limits the performance of the system.Considering the traditional prediction method can not well model the nonlinear correlation sequence and have poor performance of the long-term prediction,this thesis focuses on the downlink channel quality feedback delay problem,and uses Deep Neural Networks(DNN)to predict the downlink SINR under two kinds of different scenarios.The DNN-based SINR prediction model is proposed to mitigate the effects of channel feedback delay on the performance of wireless cellular system.The main work is as follows.1.Under parameter-invariant EPA channels,a feedforward neural network(FNN)-based prediction model for downlink SINR sequences is proposed,which realizes the SINR prediction under the scenarios of fixed UE moving speed and long channel feedback delay.By modeling and simulating the fading characteristics of the parameter-invariant EPA channel,the linear and nonlinear correlation characteristics,the overall and the short-term correlation of the SINR time series are analyzed,which is helpful to design the structural hyperparameters of the neural network.Afterthat,the grid search is used to determine the optimal structural hyperparameters.Through the threshold-based process on the neural network output,the effects of burrs are elimilated,which occurs at the neural network output when there are the peak or valley of the SINR.Compared with the traditional time series prediction model,the prediction model based on FNN can obviously improve the prediction accuracy under a variety of scenarios with fixed UE moving speed.2.Under parameter-variant EPA channels,an LSTM-FNN prediction model for downlink SINR sequences is proposed,which is based on neural network structural combination,and realizes the SINR prediction under the scenarios with dynamically changing UE moving speed and long channel feedback delay.By segmenting the SINR according to different UE speeds,the LSTM-FNN model uses the LSTM network to structurally combine multiple independent FNNs,the shortcomings of the lack of expressive ability of the FNN-based predictive model are overcome.Moreover,a pre-training&global-training strategy with an iterative fine adjustment is designed,which can lead to better convergence of LSTM-FNN model.Simulations show that the LSTM-FNN model has better predictive accuracy compared with the ARIMA model,single FNN and single LSTM neural network model,and the time complexity is within an acceptable range.