月球极低轨卫星轨道预报受高阶引力摄动影响，导致采用传统高阶引力项积分预报法通常面临极大的计算开销。针对此问题，提出了一种融合缩放点积注意力机制（Scaled Dot-Product Attention，SDPA）的改进双向长短期记忆（Bi- Long Short-Term Memory，Bi-LSTM）智能轨道预报方法。该方法使用低阶球谐函数展开式求积分，并结合Bi-LSTM模型预测其与高阶引力项积分的误差并做预测修正，能在更短计算时间内完成高精度月球极低轨预报。其中在Bi-LSTM模型中提出了具有高效的全局关键信息特征捕获能力的缩放点积注意力机制，增强了传统Bi-LSTM的学习能力和预测性能。针对15 km高度的月球极低轨任务场景，本文提出的智能轨道预报方法的最终预测误差能达到10 m量级，比传统单纯依赖积分预报的计算总耗时缩短约51.4%；改进Bi-LSTM模型相比于普通单向LSTM模型在X、Y、Z轴三方向的最终预报精度分别提升23.7%、13.8%和17.9%。所提出的具有缩放点积注意力机制的Bi-LSTM智能轨道预报能够弥补仅用低阶引力项展开求积的误差，实现在15 km高度月球极低轨环境的高精度预报，并且大幅减小了传统单纯依靠积分预测方法的时间成本。

The prediction of extremely-low lunar orbits is affected by the influence of high-order gravity perturbation, leading the traditional high-order gravity term integral method usually facing great computation cost. To address this problem, an intelligent orbit prediction method is proposed with the fusion of Scaled Dot-Product Attention (SDPA) and advanced Bi-Long Short-Term Memory (Bi-LSTM). This method uses the integration with lower-order spherical harmonic function expansion, combining the Bi- LSTM model to predict the integration error between the lower-order and the high-order gravity term and revising the prediction error, which can accomplish the high-accuracy prediction of extremely-low lunar orbits in a shorter time. Where, a SDPA with the capability of efficient global critical information feature capturing is proposed in Bi-LSTM, improving the learning ability and prediction performance of classical Bi-LSTM. For the mission scenario of extremely-low lunar orbit at 15 km altitude, the proposed intelligent orbit prediction method achieves a final prediction error of 10-meters scale, and reduces about 51.4% total computation time compared to traditional integral prediction method. The improved Bi-LSTM elevates final position prediction accuracy of 23.7%, 13.8% and 17.9% in X axis, Y axis, Z axis compared with general LSTM, respectively.The proposed advanced Bi-LSTM with SDPA intelligent orbits prediction can compensate for the error of integration which only uses the low-order gravity terms, achieving the high-precision prediction under the environment of extremely-low lunar obits with 15 km altitude. Moreover, it can greatly reduce the time cost of the traditional integration prediction method.