MSAF: A Multi-Scale Attention-Fusion Model for Long-Term Cell-Level Network Traffic Prediction

Implementing precise network traffic prediction is essential for optimizing base station strategies, allocating network resources effectively, reducing costs, and ensuring service quality in network management. Currently, most mainstream forecasting methods primarily focus on either the time domain or the frequency domain, attempting to capture and express time series data features from a one-dimensional perspective. This one-dimensional approach can be limiting, as relying solely on a single dimension may lead to incomplete or suboptimal modeling, it fails to fully capture the intricate interactions between time-domain and frequency-domain that define network traffic behaviors. To address this, we propose a novel Multi-Scale Attention-Fusion (MSAF) model. The key module in MSAF, named Attention-Fusion, integrates both time-domain and frequency-domain features of network traffic data. This effectively addresses the limitations of previous attention mechanisms in simultaneously identifying frequency components and temporal patterns, providing enhanced modeling capabilities for complex network traffic data. Additionally, the Attention-Fusion module utilizes a DSConv module combined with the attention mechanism to identify correlations between sequences of different lengths. Evaluation of MSAF on two real-world network traffic datasets demonstrates that MSAF outperforms other mainstream methods in terms of forecasting performance.