[Purposes] Accurate prediction of holiday traffic flow is essential for enhancing the operational efficiency and safety of road transportation systems. However, the complex characteristics of holiday traffic flow—including de-stationarity, nonlinearity, and abrupt fluctuations—pose significant challenges to precise forecasting. To address these challenges, a deep learning model integrating temporal convolutional network and de-stationary Transformer (TDSformer) is proposed . [Methods] In this model, TCN is first employed to capture temporal features of holiday traffic flow. Subsequently, these features are imported into a de-stationary Transformer attention mechanism, which extracts temporal dependencies directly from the raw de-stationary holiday traffic data. Finally, a typical engineering case of vehicle traffic volume during the New Year holiday on a certain expressway in Hangzhou was selected. An ablation experiment was conducted and compared with the prediction results of the existing five traffic flow prediction models. [Findings] Experimental results demonstrate that: compared to the autoregressive integrated moving average model and random forest regression, the TDSformer model reduces the mean absolute error by 64.25% and 56.64%, respectively. Compared to the long short-term memory network and gated recurrent unit , the TDSformer model reduces the mean absolute percentage error by 57.13% and 55.33%, respectively. Compared to the bidirectional gated recurrent unit-attention model, the TDSformer model reduces the root mean squared error by 48.01%. [Conclusions] The accuracy and superiority of the TDSformer model in traffic flow prediction are proved. This study provides theoretical support for optimizing traffic resource allocation and ensuring stability during large-scale holiday travel.