Fully exploiting the spatial correlation of passenger flow between related stations in the subway network has a positive effect on the improvement of subway passenger flow prediction accuracy. Capturing and quantifying spatial patterns in passenger flow data is difficult due to the difficulty of learning and transferring spatial correlations between metro stations. An improved graph-convolution gated recurrent neural network (GCGRU) metro passenger flow prediction model was proposed to enhance the model’s ability to handle different data types by integrating multivariate spatiotemporal data. The spider wasp optimisation (SWO) algorithm based on Tent chaotic mapping and Levy flight disturbance strategy was used to dynamically adjust the model structural parameters in order to optimize the hidden layer structure of the gated recurrent neural network. The experimental results show that the prediction accuracy of the model is significantly higher on weekdays than on weekends, and the root mean square error, mean absolute error, and mean absolute percentage error are reduced by 13 percentage points, 12 percentage points, and 0.08 percentage points, respectively, during weekdays compared to weekends. Dynamic optimization of the hidden structure of gated recurrent networks can lead to better convergence of the prediction model and higher prediction accuracy