Abstract: When solving physical field equations in deep learning, there are problems such as the lack of interpretability of the algorithm, the large amount of data required to train the models of the algorithm, the long training time, and the inability to modify model boundary conditions arbitrarily. To solve the above problems, the recurrent neural network (RNN) that can be used for real-time calculation of flow velocity distribution in rod bundle channels was designed in this paper. This algorithm used the RNN as its basic structure. The computational unit of the RNN was constructed using the multiple relaxation time-lattice Boltzmann method (MRT-LBM). The structure and quantity of neurons were determined using the immersion method and the method of characteristic. Then the three-dimensional model of rod bundle channels into 40 401 planes was decomposed, each plane containing 10 000 nodes. Based on the above structures, 40 401 models of RNN were established, each model including 10 000 neurons, and the velocity distribution on each two-dimensional section was calculated. Finally, the three-dimensional velocity distribution was reconstructed based on the position of the section. Due to the use of the immersion method, the method of characteristic, and the reconstruction of three-dimensional flow velocity using two-dimensional flow velocity, support vector regression (SVR) was introduced to correct the three-dimensional flow distribution. By combining the RNN and MRT-LBM, the weights and biases of RNN can be determined by the equations of MRT-LBM, and the algorithm does not require the dataset to train the models of RNN. All calculation steps of the RNN correspond to the equations of MRT-LBM, making the algorithm interpretable. Meanwhile, the MRT-LBM can utilize the parallel computing structure of RNN to further accelerate computation. The algorithm’s effectiveness was validated using the lid-driven cavity flow, the rod bundle channels, and the measurement results of PIV. The calculation results indicate that the RNN has a residual of approximately 0.1 compared to the MRT-LBM and the commercial CFD software when calculating the dimensionless velocity distribution of the two models mentioned above. The residual is one order of magnitude smaller than the flow velocity at the inlet of the model. When calculating the dimensionless flow velocity of sections in the rod bundle channels using the RNN, the calculation time consumed is about 0.005 to 0.03 seconds, which is only 1/6 to 1/3 of MRT-LBM, and the calculation results are basically consistent with the measurement results of PIV. All the computation processes of the RNN correspond to physical equations, so the RNN can greatly improve computation speed while ensuring computational accuracy, and the RNN has interpretability. The RNN can provide real-time simulation methods of flow parameters for reactor digital twin systems, further enhancing the simulation capability of digital twin systems for real-world environments.