Abstract: Abstracts: Under the international background of energy scarcity, nuclear energy has emerged as a crucial clean energy source, with proton accelerators serving as vital equipment for nuclear waste treatment. In accelerator beam orbit correction processes, accurate beam trajectory parameter prediction becomes essential for computational assistance. Given the widespread application of neural networks in industrial sectors for high-accuracy data fitting and prediction, this study proposes a beam orbit parameter prediction technique for linear accelerators based on edge intelligent computing nodes. The methodology employs a BPNN to learn historical operational data, subsequently deploying the trained model to edge acceleration nodes for predicting BPM parameters. This approach enables accelerated beam position correction through edge computing resources. Experimental results demonstrate that the FPGA hardware accelerator, functioning as an edge acceleration node, achieves a simulation inference speed of 55 μs. The energy efficiency ratio outperforms GPU and CPU platforms by factors of 23.13 and 553.15 respectively, with an average prediction error of 0.5%. Both temporal performance and prediction accuracy meet predetermined requirements, indicating promising potential for real-time beam control applications in nuclear energy facilities.