5G Network Traffic Prediction Based on WP - Deep Gaussian

Abstract

Network traffic is nonlinear and nonsmooth, so it is difficult to accurately predict long-term traffic. To improve the accuracy of network traffic prediction, this paper proposes a WP-depth Gaussian network traffic prediction model using the wavelet denoising method and deep Gaussian process. Firstly, the traffic sequences containing noisy signals are decomposed using wavelet basis functions, and the high-frequency signal sequences are thresholded to remove the noise. In addition, the denoised signal sequence is output through the inverse wavelet transform to complete the noise reduction of the data. Finally, the variational inference is introduced in the posterior inference. The posterior probabilities are optimized by using a deep Gaussian process to obtain the optimal hyperparameters of the kernel functions of each layer. While completing the uncertainty measurement of the predicted values, the network traffic prediction results with high accuracy are given. Finally, BP neural network, LSTM neural network, and ARIMA time prediction model are constructed as comparison models. In this experiment, three error analysis metrics, MAPE, MAE, and RMSE, are used to describe the prediction effects of the four models. The experiments show that the three prediction error ratings of the WP-Depth Gaussian prediction model are smaller than the error metrics of the other three models, and have sound prediction effect. It shows that the Deep Gaussian prediction model combined with wavelet denoising is reasonable and feasible for network traffic prediction.