Abstract:

The rapid growth of passenger volume of urban rail transit and the urgent need for intelligent operation have made accurate section passenger flow prediction a key technical challenge to improve the level of dynamic scheduling and safety control. To this end, this paper innovatively integrated the physical information constraint mechanism, data-driven method and SGA, and proposed a new deep learning framework. Firstly, a physical residual term was designed and embedded into memory cells as a regulation signal, forcing the model to learn the physical laws of passenger flow while retaining the temporal characteristics of passenger flow. Secondly, a dual-objective fitness function based on physical loss and data loss was innovatively proposed to achieve further optimization of model performance while establishing a constraint mechanism. Finally, SGA was used to balance the differentiated and synergistic effects of hyperparameters in the model. Experimental results show that the constructed model exhibits good predictive performance on both the training set and the validation set. The improved fitness function can narrow the error range of the model prediction results. In the two-stage ablation experiment, the mean square error range of the proposed deep learning framework is reduced by 71.03% compared with the long short-term memory model, which verifies the synergistic enhancement of the model prediction ability by the simultaneous introduction of physical constraint mechanism and intelligent optimization algorithm.

Key words: physical information, snow geese algorithm (SGA), subway, section passenger flow, deep learning