Multi-Timescale Feature Fusion Enhanced Short-Term Network Traffic Burst Prediction

Accurate modeling and prediction of network traffic are crucial in wireless networks. However, network traffic bursts occur with high randomness in timing and magnitude. Traditional methods primarily analyze long-term trends and cyclic patterns of traffic, often resulting in inaccurate prediction of short-term traffic bursts and significant errors. Consequently, accurately forecasting fine-grained network traffic poses a challenge. To tackle this issue, we propose a novel traffic prediction model. Specifically, to capture the burstiness of the traffic, we design a CNN-GRU structure to capture rapid changes in traffic values over local short time intervals. Additionally, to account for the long-term periodic characteristics of traffic, we apply Fast Fourier Transform (FFT) to the traffic data and extract the main frequency components. To combine both burstiness and periodicity in traffic patterns, we employ a fusion network to capture multi-timescale feature. This approach enables the model to interpret traffic data from various perspectives, enhancing its adaptability to changes in traffic patterns due to bursts and ensuring robust predictive performance. Experimental results demonstrate that our framework outperforms the baselines, achieving a 14.5% reduction in prediction error for burst traffic on the China Unicom One Cell dataset.