A Moment Cross Predictor For Non-stationary Mobile Traffic Forecasting

Mobile traffic forecasting plays a pivotal role in optimizing network resources and ensuring Quality-of-Service (QoS). Accurate traffic forecasting enables network operators to configure the network proactively to meet user requirements. However, real-world mobile traffic is non-stationary, denoting traffic distribution changes over time. Most prediction models are based on the assumption of stationarity. Therefore, improving non-stationary mobile traffic forecasting poses a significant challenge due to the varying statistic distribution and the varying moments. We propose the Moment Cross Predictor (MCP) to solve this challenge. Firstly, we propose the high-order moment normalization to model the complex time-varying distribution. Secondly, we use the Koopman operator to capture the dynamic evolution of each order moment and use a multi-head self-attention mechanism to model the interdependence among moments. Consequently, MCP can serve as a plugin to effectively decouple the non-stationarity nature of mobile traffic. We instantiate MCP with four forecasting models on mobile network traffic datasets. The results show that the model with MCP outperforms baseline models on evaluation metrics.