Long-term network traffic prediction based on learnable trend seasonal decomposition

The continuous growth of network traffic has introduced increasingly complex challenges in network management. Accurate forecasting of future network traffic is essential for optimizing network utilization. Mobile operators can leverage long-term network traffic forecasts to optimize network planning and resource allocation, thereby enhancing the quality of service provided to users. Current network traffic prediction methods either utilize moving average (MA) techniques to decompose input network traffic into trends or employ learnable network layers to achieve decomposition into trend and seasonal components. However, the selection of convolution kernels in these methods often necessitates extensive experimentation, making the prediction performance highly dependent on prior knowledge. To overcome these limitations, we propose a neural network model incorporating learnable trend-seasonal decomposition for longterm mobile network traffic prediction. To achieve time series decomposition, we replace the single fixed convolution kernel with multiple learnable convolutional layers. Furthermore, we adopt a down-convolution-up-convolution network architecture to effectively extract the traffic trends, minimizing information loss during the convolution process. In the experiments, we evaluated the predictive performance of the proposed model in comparison with the current state-of-theart models. The experimental results demonstrate that, compared to various existing TSP(Time Series Forecasting) prediction models, the proposed model achieves superior accuracy in longterm mobile network traffic forecasting, underscoring its potential for effective application in long-term traffic analysis.