A Study Of Neutral Network Application In Pilot Based OFMDA Channel Prediction

Orthogonal Frequency Division Multiple Access (OFDMA) is the downlink channel technology in next generation wireless communication system LTE. Dynamic bit loading adaptation link technology can be used to maximize OFDMA channel capacity. Link adaptation algorithms have to get the coming channel state which is usually unknown to the transmitter. However, there exists feedback delay error in estimated channel states and the error always gets worse in fast fading channels.Channel prediction methods can be used to predict the coming channel state information in advance, so it can reduce the impact of channel feedback delay error. Neutral network can also be used to predict channel state so as to provide link adaption with more accurate channel state information.This paper deploys neutral network algorithms to perform pilot based LTE OFDMA channel status prediction, and advances a channel state prediction model based on neutral networks. The prediction model takes history pilot channel states as neutral network input to predict the coming pilot channel state information, and the prediction model output is derived by adjusting neutral network prediction outputs based on the average pilot channel state information and the average square pilot channel prediction error in a the past period. In this paper's prediction model, neutral network can use LM network or RBF network. The following method is used, at the receiver part, to evaluate the prediction efficiency: firstly, during simulation, use the prediction model's output as the LTE link level simulation system's input, then generate the simulation system's Bit Error Rate (BER) and Block Error Rate (BLER), at last, compare the simulation result with those without prediction. The simulation results of this paper's prediction model is: in a certain prediction steps, the predicted channel state information is very close to the real channel state, and the link level BER and BLER are also close to those without prediction. In addition, different neutral networks in the prediction model provide different link level performance data while requiring different number neurons, network training duration and network prediction duration. So these factors have to be considered before deployment.