Research On Spectrum Prediction Algorithm Based On LSTM

Spectrum scarcity has become a major problem to be solved urgently in the era of wireless communication.In addition to the objective factor of limited spectrum resources,the most fundamental cause of this phenomenon is actually the low utilization efficiency of spectrum in the static allocation of resources mode,which has been proved by enough spectrum measurement results.The cognitive radio technology is regarded as a powerful means to improve the spectrum efficiency,mainly including spectrum sensing,spectrum decision,spectrum sharing and spectrum switching,but with the deepening of the research,people gradually find that there are many problems with these four technologies,such as vast energy consumption,communication time delay and communication conflict.Therefore,people put forward the spectrum prediction technology to assist cognition,which can effectively alleviate the above problems and improve the accuracy of finding and accessing idle frequency bands.Based on this,the spectrum prediction algorithm is deeply studied in this paperFirstly,the existing spectrum occupancy models are analysed and researched,and according to the summary results,queuing theory is selected as the foundation simulation model for this paper due to its high frequency of use,then combining the principle of the influence in the practical wireless communication system,a spectrum occupancy model based on allocation rules and queuing theory is proposed,which takes into account the spectrum usage of multiple channels.Besides,two classical traditional spectrum prediction algorithms are introduced and simulated on the data set generated by the above determined spectrum occupancy model.Next,the single-step spectrum prediction algorithm is studied.LSTM is used to forecast the single channel spectrum usage out of consideration for the correlation of spectrum data in time,and in order not to be limited by the one-way structure,Bi-LSTM is chose to extract the two-way correlation.The above two algorithms are simulated and compared,and some errors in the actual data are also considered.The results show that the prediction performance of Bi-LSTM is better than that of LSTM,and both of them are better than the traditional prediction algorithms.In addition,considering the correlation of spectrum data in time and between channels,multiple channels are selected to predict.ConvLSTM with convolution calculation is adopted and Bi-convLSTM is proposed with reference to Bi-LSTM for simulation comparison.The results show that the prediction accuracy of Bi-convLSTM is higher than that of convLSTM,and the algorithm based Bi-convLSTM is less affected by wrong data,that means the algorithm performance is more stable.Compared with single-step prediction,multi-step prediction can obtain more channel usage at one time,however,multi-step prediction has the problem that the prediction accuracy will drop significantly with the increase of the prediction step length.In this paper,Seq-to-Seq model and LSTM are combined to propose an algorithm for multi-step spectrum prediction by combining multiple channels,and attention mechanism is introduced to further optimize the algorithm.The simulation results show that the performance of the multi-channel multi-step spectrum prediction algorithm with the attention mechanism is better than that without the attention mechanism,and they are both better than that of general single-channel multi-step prediction algorithm.The prediction accuracy of above algorithm decreases with error rate but remained within the acceptable range.