基于长短时记忆神经网络(LSTM)的钟差预报算法

原子钟钟差预报在原子时计算和原子钟频率驾驭中发挥着重要的作用。长短时记忆神经网络（LSTM）预报算法能够处理多参数长期依赖关系的时间序列预报，以氢钟和铯钟实测数据为样本，通过构建LSTM钟差预报模型，降低了长期原子钟内部噪声以及原子钟漂移对钟差预报的影响，并以72h，240h和720h为预报时长，分别与线性多项式模型、灰色模型和Kalman模型原子钟钟差预报模型进行预报误差对比。研究表明，在240h以上的预报时长中，LSTM建模长期依赖关系的优势得以体现，相较于其他3类模型可以获得更高的预报精度。

Long Short Term Memory Neural Network (LSTM) based on Clock Difference Forecasting Algorithm

Atomic clock clock difference forecasting has an important role in atomic time calculation and atomic clock frequency harnessing.In this paper,we use hydrogen and cesium clocks as samples to reduce the effects of long-term atomic clock internal noise and atomic clock drift on the clock difference prediction by constructing LSTM clock difference prediction models,and compare the prediction lengths of 72 h,240 h and 720h with linear polynomial model,gray model and Kalman model respectively.The prediction errors are compared with the linear polynomial model,the gray model and the Kalman model.The analysis shows that for forecast lengths above 240 h,the advantage of long-term dependence of LSTM modeling is realized,and higher forecast accuracy can be obtained compared with the other three types of models.