Cellular Traffic Prediction Based on Two-Stage Cross-City Transfer Learning

The imbalance in network resource allocation and service quality across cities remains a critical issue that demands effective solutions. Accurate city-scale cellular traffic prediction plays a vital role in addressing this challenge. However, existing methods heavily rely on abundant data, limiting their applicability in data-scarce cities. To overcome this limitation, we propose a two-stage Cross-city Transfer Learning framework for cellular traffic Prediction (CTLP). In the first stage, a spatiotemporal feature concatenation network is proposed. This network captures the dynamics of cellular traffic derived from a data-rich city and a data-scarce city with attention mechanisms and then aggregates these dynamics with a CNN. In the second stage, parameters of the first-stage network are transferred from the data-rich cities to the data-scarce cities for addressing the challenge of cellular traffic heterogeneity in cross-city transfer and enabling more accurate predictions. Experiments on a realworld cross-city cellular traffic dataset demonstrate that CTLP significantly outperforms existing methods, effectively solving the problem of cellular traffic prediction in data-scarce cities.