Abstract: An in-depth analysis of the limitations of traditional time series models reveals that they mainly rely on linear assumptions, which makes it difficult to effectively capture the nonlinear dynamic relationships in complex data. To address this problem, a dual-path design incorporating fully connected layers and KAN networks is investigated, and a BiLSTM-MCEP hybrid model is constructed for short-term traffic flow prediction based on the BiLSTM model, incorporating multiscale convolution and efficient channel attention mechanism. By adopting an end-to-end learning framework model and utilizing the advantageous complementary properties among neural networks to fully mine historical data from different time periods, thus improving the overall performance and prediction efficiency of the model. Based on the PeMS traffic dataset, the experimental validation is carried out, and the results show that the model has an RMSE of 0.04815, an MAE of 0.03543, an MSE of 0.00232, and an R² of 0.94368. Comparison of the experimental results shows that, compared with the BiLSTM model, the method achieves a significant increase in modelling accuracy and robustness, and demonstrates strong prediction ability.