Abstract: A more accurate prediction of subway passenger flow requires the integration of long- term and short- term characteristics to fully capture both periodic and non- periodic features. To tackle this issue, we propose a metro passenger flow prediction model called IbCNNM- MSTL- TFT, which combines learning and non- learning methods. The model analyzed the internal regularity and periodicity of the time- series data across multiple time scales, replacing the single time scale used in traditional models, which allowed for the efficient integration of multiple features and the generation of prediction results. Experimental results from 15 stations on a subway line indicate that the IbCNNM- MSTL- TFT prediction error is significantly lower compared with the benchmark model TFT at 11 stations. The mean absolute error (MAE) was decreased by 0.9, and the root mean square error (RMSE) was decreased by 3.7. The model's prediction accuracy surpasses that of many other current deep- learning prediction models, and it demonstrates strong predictive capability.