Precise passenger flow forecasting is the basis for rail transit planning. Based on the dynamic, nonlinear, uncertain, periodic, non-stationary and sequential characteristics of urban rail transit’s short-term passenger flow, this paper proposes a combined model prediction method to improve the prediction accuracy of short-term passenger flow, namely VMD-GRU neural network prediction model. This model is composed of variational mode decomposition and gated recurrent units to realize forecast of short-term passenger flow of urban rail transit. According to the data of Nanjing Metro, the model has a good effect on the short-term passenger flow prediction, and is superior to single models such as SVM, BP, GRU, and hybrid models such as VMD-SVM and VMD-BP. It can provide effective data support for subway operation management departments in terms of passenger flow management at stations and the formulation of daily driving plans, thereby improving the overall operating efficiency of the line network and the service level of the rail transit system.