| The Earth rotation parameters(ERP):px,py pole coordinates, universal time(UT1-UTC) and length-of-day change((35)LOD), describe the variations in Earth rotation. The near real-time estimates of ERP are required for various fields linked to reference systems such as interplanetary spacecraft tracking and navigation, satellite navigation and astrogeodynamics. ERP estimates enable the time-varying transformation between the celestial and terrestrial reference frames(CRF and TRF).Owing to the complicated process of data processing, the ERP estimates obtained by advanced space geodetic technologies delay by several days to two weeks in general.The need to provide better ERP forecasts for space navigation has promoted more activity in high accuracy prediction of ERP. Some investigations are carried out to enhance the prediction accuracy of ERP in this thesis, e.g., the improvement of the edge-effect of least squares(LS) fitting, the influence of dataset size length on the prediction accuracy of ERP, the combined solution for ERP prediction. The detailed work can be described as follows.(1) For the purpose of eliminating the edge-effect of LS fitting, the time-series analysis model is first used to extrapolate ERP time-series forward and backward and then generate a new time-series. Subsequently, the coefficients of a LS model are estimated by using the new generated time-series. As a result, the edge-effect is changed to the edge of the new time-series, and thus the original fitted time-series can be free of the edge-effect. Finally, the combination of LS and autoregressive(AR)models(LS+AR) is employed to predict the original ERP. The results of simulation experiments and an application to ERP prediction show that the proposed method,i.e., the edge-effect corrected LS(ECLS) model can efficiently eliminate the edge-effect, and thus improve the prediction accuracy of the LS model, especially for medium- and long-term prediction.(2) On one hand, the metabolic, interval and recursive patterns for AR filtering are analyzed and compared. The results indicate that the prediction accuracy of the metabolic pattern is worst with the lowest efficiency out of the three patterns, and both the prediction accuracy and efficiency of the recursive pattern are highest, while these of the interval pattern are mediate among them. On the other hand, a simple differencing technology is employed to enhance the prediction equality of UT1-UTC.This technology first makes one-order difference between adjacent epochs of UT1-UTC time-series and then the differenced sequence is obtained. Next the differenced sequence is modeled and predicted by the combination of LS and AR models. Ultimately, the predictions of the differenced sequence are recovered to yield the UT1-UTC forecasts. It is demonstrated that the differenced time-series are more stationary and thus is high compatible with an AR model. In addition, the prediction quality can be substantially enhanced with the proposed differencing technology.(3) The influence of dataset size on the prediction accuracy of ERP is analyzed and discussed. It is found that the forecasted results are strongly associated with dataset size. Generally speaking, the more the dataset size is, the better the prediction accuracy of(35)LOD and UT1-UTC is, but it is clean contrary in the case of polar motion. Therefore, the dataset size should be carefully taken into account in practical work.(4) The ERP are predicted by means of the neural network(NN) technology.The algorithm for network topology design is first researched. A minimum comprehensive root mean square error(RMSE)-based algorithm is presented for determining network topology. Using this algorithm the training and testing error is together considered in the strategy, and thus the over-fitting and under-fitting can be avoided. Then the NN prediction algorithm procedure is described in detail and the training patterns for NN prediction are composed, including the continual, interval and recursive patterns. It is found that the short-term prediction accuracy of interval and recursive patterns is high, while the long-term predictions obtained by the continual pattern are better. As for the computational efficiency, the recursive pattern occupies first place among them, the interval pattern comes second, and another comes third. Finally, the prediction results obtained by the one hidden-layer,multi-input and single-out NN model are analyzed and compared with those of the Earth Orientation Parameters Prediction Comparison Campaign(EOP PCC). The results illustrate that the prediction accuracy of the developed method is close to the international advanced level.(5) An EMD+NN-based extrapolation approach is proposed to reduce the contribution of high frequency signals in pole coordinates data to prediction error,where the empirical mode decomposition(EMD) algorithm is used for removing high frequency signals in polar coordinates data. The performance of this system,trained using the EMD+NN method on the preprocessed time-series with high frequency signals removed, is compared with that of a NN procedure trained on raw time-series. The results manifest that the predictions by the EMD+NN method are superior to the NN-only model for medium-term(up to 90 days in future). However,the improvement of prediction accuracy is found to be decreased with increase of prediction time.(6) A grey neural network(GNN) model created from(35)LOD and UT1-UTC time-series, where a grey layer and a white layer are connected to input and out layers, respectively, is summarized. In the grey layer(35)LOD and UT1-UTC residuals are preprocessed by the Inverse Accumulated Generating Operation(AGO) to alleviate non-stationarity and persistent autocorrelations, while NN outputs are transformed with the Accumulated Generating Operation(AGO) in the white layer.This method has advantage over the NN-only approach that can be easily influenced on by non-stationarity and persistent autocorrelations. It is demonstrated that the GNN model noticeably outperforms the NN-only solutions, in particular for medium- and long-term prediction.(7) A main conclusion of the EOP PCC is that there is not one particular prediction technique superior to the others for all EOP and all prediction intervals,and the combined prediction is shown to perform very well as do some of the individual prediction. In consideration of the advantages of using a combined solution, an ensemble prediction method is developed simply based on the weighted mean of individual prediction available at a given prediction epoch, where the weights assigned to ERP forecasts are calculated according to the standard deviation and RMSE of prediction residuals. The merits of utilizing a combined solution are clearly demonstrated as the ensemble method very often performs better than all individual prediction techniques and is quite naturally more reliable and robust with respect to mistakes or human errors. The main conclusion is that the prediction accuracy of ERP can be remarkably improved by using the proposed ensemble prediction solution.(8) The ERP prediction software is compiled by the National Time Service Center(NTSC) in order to yield daily routine ERP predictions. The results of ERP predictions are compared with those obtained by the International Earth Rotation and Reference Systems Service(IERS) product center. It is demonstrated that the NTSC products are more accurate than those published by the IERS product centerespecially for medium- and long-term prediction. |
PDF Full Text Request