Abstract: Since accelerator beam orbit correction is crucial for stable operation of accelerators, accurate prediction of the changes of the accelerator beam orbit is also essential for automated beam calibration. Precise predictions of beam orbit changes can provide reliable information for adjusting accelerator control parameters to achieve precise control and regulation of beam orbit. In this paper, based on a multi-stage cascaded back propagation (BP) neural network and simulated accelerator data, the beam transport process in the medium energy transfer section of a linear accelerator is studied and an accelerator beam orbit prediction model is constructed to predict beam orbit parameters. The results show that the proposed multi-stage cascaded BP neural network achieves higher prediction accuracy and reliability than the prediction model built using a traditional single hidden layer BP neural network. This provides an effective method for optimally designing the medium energy transfer section of the linear accelerator and automating the calibration of the beam orbit.