Research On Spectrum Prediction Based On Recurrent Neural Network

With the increase of mobile devices,the available radio spectrum resources are limited,and there is an imbalance between the supply and demand of spectrum resources.However,the spectrum utilization rate of the previous fixed spectrum allocation strategy is low,which cannot meet the increasing frequency devices,resulting in a lot of waste of spectrum resources.Spectrum prediction technology is an effective method to improve spectrum utilization.With the development of deep learning,the technology of neural network becomes more mature,and it is feasible to apply deep learning technology to spectrum prediction.This paper will focus on the spectrum prediction technology based on recurrent neural network.In cognitive radio,spectrum prediction is a very important research content.Spectrum prediction includes many aspects,including spatial domain spectrum prediction,time-frequency domain energy prediction,and channel state prediction.The research content of this paper has two points,one is spatial domain spectrum prediction and the other is channel state prediction.In spatial domain spectrum prediction,the main research question is how to predict the signal strength of the entire space in a given attenuated occlusion space,where the signal strength measurements at the locations of known finite points are predicted.If you are in a simple environment,you can use mathematical or physical formulas.Three-point positioning is performed to infer the position of the emitter,thereby deriving the signal strength values of other grids.This article considers a more complex problem where the transmitter location is unknown and the space environment is complex.In this paper,combined with the spatial characteristics of the data,a special CSRN network is used to solve the problem,and compared with the MLP method.In the study of channel state prediction,the main research question in this paper is how to use the historical channel state value in a certain period to predict the channel state value in a certain time slot in the future.In this paper,the correlation characteristics of channel state in time are fully considered.The seq-to-seq network structure based on recurrent neural network has the characteristics of long-term and short-term memory function,which is applied to this problem.It is studied how to use the channel state values of n historical time slots to predict the future m channel state values.Moreover,the influence of the input sequence length n on the prediction accuracy,and the effect of the predicted slot length m on the performance is discussed.Moreover,with the seq-to-seq network structure,simultaneous joint prediction for multiple users is also possible.In addition,this paper discusses the performance difference between multi-user joint prediction and single user independent prediction.The experimental results in this paper show that the spatial spectrum prediction technology based on CSRN network proposed in this paper can solve the problem of spatial spectrum prediction in complex environment with unknown emitter position and unknown environmental information,without retraining the network,to achieve better performance than MLP performance.Moreover,the larger the parameter T,the better the performance of the CSRN network.In addition,the channel state prediction based on the seq-to-seq network proposed in this paper can learn the rules existing in the input sequence,and the performance is obviously better than the MLP.and It is also much better than the independent prediction performance in the joint prediction of multiple users.