Abstract
This paper proposes a hybrid model combining Transformer and Temporal Convolutional Network (TCN). This model addresses the shortcomings of current approaches in capturing long-term and short-term dependencies in network traffic prediction tasks. The Transformer module effectively captures global temporal relationships through a multi-head self-attention mechanism. Meanwhile, the TCN module models local and long-term dependencies using dilated convolution technology. Experimental results on the PeMSD4 and PeMSD8 datasets demonstrate that our method considerably surpasses current mainstream methods at all time steps, particularly in long-term step prediction. Through ablation experiments, we verified the contribution of each module in the model to the performance, further proving the key role of the Transformer and TCN modules in improving prediction performance.