利用小波分解改进极移预报模型

高精度的日长（length of day，LOD）变化ΔLOD预报值在深空探测器跟踪、卫星自主导航和气候预测等领域具有重要作用。针对ΔLOD复杂的时变特性，首先，利用奇异谱分析（singular spectrum analysis，SSA）方法提取ΔLOD时间序列中的趋势项、周年项与半周年项等主成分，并基于SSA迭代插值算法对主成分进行外推；其次，采用差分自回归滑动平均（autoregressive integrated moving average，ARIMA）模型对扣除主成分的剩余项进行建模预测；最后，将SSA主成分外推值与ARIMA预测值相加获得ΔLOD预报值。选取国际地球自转与参考系服务组织发布的2000-01-01—2001-12-31的ΔLOD数据进行1~365 d跨度的预报实验，并将SSA+ARIMA预报结果与反向传播神经网络、广义回归神经网络和高斯过程等机器学习方法的预报结果进行对比分析。结果表明，SSA+ARIMA方法的预报精度优于几种机器学习方法，特别是中长期预报精度优势更为显著，其中，对于1~30 d短期和30~365 d中长期的预报，SSA+ARIMA方法的平均绝对预报误差相对于机器学习方法最大分别降低了39%和61%。

Earth Rotation Monitoring,UT1 Determination and Prediction

Length of day variation (ΔLOD) predictions play an important role in tracking and navigation of deep-space detector, precise determination of artificial satellite orbit and climate forecasting. In full consideration of the time-varying characteristics of ΔLOD, we present the application of a hybrid technique for predicting ΔLOD.The ΔLOD predictions are generated by means of the combination of singular spectrum analysis (SSA) extrapolation for the linear trend, annual and semiannual oscillations in ΔLOD based on an iterative interpolation strategy, and autoregressive integrated moving average (ARIMA) stochastic prediction of SSA remaining residuals, referred to as SSA+ARIMA. In order to evaluate the effectiveness of this approach, the ΔLOD predictions up to 365 d into the future are calculated year-by-year for the 2-year period from Jan. 1, 2000 to Dec. 31, 2001 using the data covering the previous 10 years from International Earth Rotation and Reference Systems Service C04 series. The prediction results are analyzed and compared with those obtained by machine learning methods such as back propagation neural network (BPNN), general regression neural network (GRNN) and Gaussian process (GP).It is shown that the accuracy of the predictions are better than that by machine learning methods in terms of the mean absolute error (MAE) of predictions, especially for medium and long-term predictions. Compared with the predictions obtained by the BPNN, GRNN and GP, the MAE of the proposed SSA+ARIMA predictions up to 30 d and 365 d in future is reduced by 39% and 61%, respectively.The proposed method provides a new solution for ΔLOD prediction.