Research On Spectrum Occupancy Prediction Method Based On Recurrent Neural Network

In recent years,with the continuous development of the information society,modern society is more and more dependent on the instant communication technology,wireless communication is widely used as a convenient means of communication.However,the frequency resources suitable for wireless communication are limited.Therefore,it is an urgent need for modern information society to divide communication channels reasonably and improve the use efficiency of spectrum resources by technical means.Spectrum occupancy is an important index to measure whether the spectrum is used effectively,In this paper,spectrum occupancy prediction is studied.Two different research methods are proposed for short-term prediction and long-term prediction.In order to improve the accuracy of prediction,different noise reduction algorithms are used to process spectrum occupancy information.EMD-LSTM network is proposed for short-term prediction of spectrum occupancy,and new Res-Ind RNN network is proposed for long-term prediction.The main work is as follows:(1)In the short-term prediction of spectrum occupancy,firstly,the timing sequence network LSTM is more suitable for the prediction of spectrum occupancy;secondly,the EMD decomposition algorithm is used to extract feature sequences of various scales and fuse the original sequences to improve the short-term prediction accuracy of the neural network;finally,the EMD-LSTM network is constructed for the prediction.The experimental results show that compared with the traditional stationary time series modeling method and the common neural network,it has a better prediction effect.(2)In the long-term prediction of spectrum occupancy,the problem of excessive length and complexity of sequence is put forward,and the network needs to be deepened to improve the prediction accuracy.On this basis,wavelet transform,a more suitable data processing algorithm for long-term prediction,is adopted to perform data noise reduction.The Ind RNN network was selected to improve its internal structure,and the in-batch normalization method of internal components was replaced with the layer normalization method to make it more suitable for dealing with time series problems.Meanwhile,the new Res-Ind RNN neural network was constructed to carry out experiments based on residual network technology.Experimental results show that the performance of the new algorithm is better than that of conventional neural network(BPNN,RNN,LSTM)and Res-Ind RNN neural network without internal structure improvement.