The high nonlinearity, strong temporal dependencies, feature redundancy, and noise in traffic flow data reduce the prediction accuracy of models. To address the challenges, a short-term traffic flow prediction algorithm was proposed. The algorithm integrated complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and a soft thresholding temporal convolutional network (STTCN). Historical traffic flow data were decomposed into high-frequency and low-frequency components using the CEEMDAN algorithm. A timestamp encoding was designed to process time information, and the maximum information coefficient (MIC) was used to examine the correlation of time and weather features with the two decomposed components. The most relevant features and their corresponding high and low-frequency components were input into the STTCN. A soft thresholding parameter, adjusted by the slime mould algorithm (SMA), was incorporated to enhance the network’s ability to handle high-noise data. The high and low-frequency components were reconstructed into traffic flow predictions. In a Zhejiang highway dataset, the proposed algorithm significantly outperforms baseline models, with reductions of 54.97% in mean squared error, 30.07% in root mean square error, and 34.39% in absolute deviation. Results show the proposed algorithm’s effectiveness in capturing the intricate dynamics of traffic flow.