基于噪声统计的EMD-LSTM网络流量预测方法

Method of an EMD-LSTM Network Traffic Prediction Based on Noise Statistics

为了提高LSTM对网络流量的预测精度,针对训练样本中存在噪声的问题提出了一种基于噪声统计特性的EMD降噪方法,将每一训练样本经EMD分解后得到若干IMF,通过分析这些IMF确定其中的噪声IMF,将各样本中同一位置的噪声IMF进行统计平均后再与每一样本中的非噪声IMF叠加,以此实现每一样本的降噪,并将降噪后的样本作为LSTM训练样本;针对LSTM中使用滑动窗口输入训练样本时存在的误差叠加问题使用间隔采样的输入方法构造训练样本;将两种方法结合提出一种EMD-LSTM预测模型;仿真表明,相较于传统LSTM预测模型,所提模型具备更优的降噪效果与更准确的预测结果;此外,所提预测模型应用于一种基于无人机卸载流量的蜂窝网络,基于该模型预测结果提出一种无人机活动规划方法以优化无人机长时间工作中返航充电的时间点,规划后的无人机在同等情况下可以使用更小的缓存队列应对突发流量。

In order to improve the prediction accuracy of long short term memory(LSTM)for network traffic,aiming at the problem of noise in training sample,an empirical mode decomposition(EMD)noise reduction method based on the noise statistical characteristics is proposed.Firstly,each training sample is decomposed by the EMD to obtain several intrinsic mode functions(IMFs),and the noise IMFs are determined by analyzing these IMFs.The noise IMFs at the same position in each sample are statistically averaged and then superimposed with the non-noise IMFs in each sample,the noise reduction of each sample is realized by this way,and the denoised sample is used as the LSTM training sample;Secondly,aiming at the error superposition problem,when the training sample is inputted by using sliding window in LSTM,the training sample is constructed by using the input method of interval sampling.Combining with two methods,an EMD-LSTM prediction model is proposed.The simulation shows that compared with the traditional LSTM prediction model,the proposed model has better noise reduction effect and more accurate prediction result.In addition,the prediction model is applied to the cellular network by using UAV to the offload traffic,based on the prediction results of the model,a UAV activity planning method is proposed to optimize the time of the UAV returning to home for charging during long-term work.This planning method can use the smaller buffer queue to deal with burst traffic under the same conditions.