Abstract: To address the issue that traditional prediction models in Software-Defined Networking （SDN） struggle to capture long-term correlations in traffic sequences and exhibit low accuracy in multi-step prediction tasks, this paper proposes a multi-step network traffic prediction model fusing Convolutional Neural Network （CNN）, Gated Recurrent Unit （GRU）, and ATTENTION mechanism. The model uses CNN to extract spatiotemporal features of traffic data, leverages GRU to adaptively learn temporal patterns, and employs ATTENTION mechanism to assign weights, alleviating information decay in long sequences for multi-step prediction. Multi-step prediction experiments were conducted by collecting SDN network traffic in a Mininet simulation environment, initially verifying the model's effectiveness. Further testing on the Telecom Italia real-world dataset shows that the model achieves at least 16.05%, 21.10%, and 3.35% performance improvements over CNN-GRU, CNN-LSTM and other mainstream models in MAE, MAPE, and R² metrics, respectively. This demonstrates the model's superiority in mining long-term correlation features, providing a higher-precision solution for multi-step prediction of SDN network traffic.