Satellite orbital state is complex and highly variable due to the influence of space environment. Existing dynamics models struggle to fully and accurately describe the effects of spatial perturbations. To address the issue of decreasing orbit accuracy over time due to insufficient dynamics modeling, this paper proposes a novel orbit prediction method based on Bidirectional Long Short-Term Memory (BiLSTM) network. BiLSTM learns the time-domain features of orbit prediction errors and then predicts future errors to modify the orbit. This method effectively compensates for the increasing discrepancy between idealized model and real dynamics environment, mitigating rapid divergence of errors. Using the method proposed in this paper, the 7-day orbit data of BeiDou MEO PC22 satellite was tested. In J2000 coordinate system, the maximum error of three-dimensional position decreases from 8395 m to 261 m, achieving a maximum error improvement rate of 96.89%. This algorithm outperforms BP network and LSTM network, which achieve improvement rate of 80.35% and 96.61% respectively.