Empirical Mode Decomposition and Stationary Wavelet Transformation in Internet Traffic Prediction

Optimizing the usage of network resources relies heavily on accurate and efficient internet traffic forecasting. One commonly used method for traffic forecasting involves decomposing the traffic signal into hierarchical components using Empirical Mode Decomposition (EMD) and using different algorithms to predict each component individually. However, this approach can be inefficient, as the number of components varies depending on the input signal, making it challenging to propose and optimize separate prediction models for each component. To address this issue, we propose a new method that utilizes a multi-output single prediction model to forecast each EMD component of target traffic volume individually and combines the individual predictions for the final forecast. We also compared the performance of our EMD-based approach with a Stationary Wavelet Transformation (SWT) integrated deep learning model. We used SWT to decompose our real-world internet traffic into high and low-frequency components, which respectively represent the white noise and long-term trend in the traffic data. To train our deep learning models, we only considered the low-frequency component to extract the trend feature. Our experimental results showed that the EMD-integrated model outperformed traditional deep learning models by approximately 1% to 3%. Furthermore, we improved the traffic prediction accuracy by smoothing the original traffic based on the SWT denoising process. This method outperformed the baseline deep learning model by an additional 3% −5%. Overall, our proposed method offers a more efficient and accurate approach for internet traffic forecasting, and the SWT-based denoising process further enhances its performance.