Studies On The Prediction Of Earth's Variable Rotation By Artificial Neural Networks

Prediction of the variations of the Earth's variable rotation is of great scientificand practical importance. However, due to the complicated time-variablecharacteristics of the Earth's variable rotation, it's usually difficult to obtainsatisfied predictions by conventional linear time series analysis methods. This studyemploys the non-linear artificial neural networks (ANN) to predict the variations ofthe Earth's rotation.As the solid Earth and its surrounding fluid layers form an approximately closedynamic system, changes of atmospheric or oceanic angular momentum will resultin variations in the solid Earth's rotation, based on the law of conversation ofangular momentum. The high-accuracy Earth rotation observations and researcheson global atmospheric models reveal that the axial atmospheric angular momentum(AAM)χ₃ correlates strongly with the LOD changes and the equatorial AAM 2χ₁andχ₂ correlate with the polar motion excitation. Thus, when the AAM series isincorporated into the prediction of the Earth's variable rotation, it will add aphysical constraint to the prediction. The present study focuses on incorporating theAAM series into the prediction of the Earth's variable rotation to improveaccuracies of the ERP predictions by ANN. In addition, the technology is alsoapplied to predict the E1 Nino/Southern Oscillation (ENSO) event.The main research work of this thesis can be summarized as follows:(1) The algorithms of determining the topology of an ANN is analyzed, and theRoot Mean Squared Error (RMSE) is chosen as the criterion to determine thetopology of the network. The network algorithm flow that is suitable for our work is investigated.(2) Based on the ANN technique, the variations of the Earth's rotational rate (i.e.,length of day, LOD) are predicted in three ways, i.e., using LODR only, using bothLODR and AAM data, and real-time rapid approach, respectively, a) The resultsshow that ANN has effective non-linear prediction ability, b) As the atmosphere isthe main excitation source of the LOD change, the accuracies of predictions aresignificantly improved after introducing the AAM into the LOD prediction,especially for the long prediction intervals, c) Real-time rapid prediction is of greatscientific and practical importance. In this thesis we introduce the operationalprediction series ofχ₃, which is from National Centers for EnvironmentalPrediction (NCEP), to the prediction set of ANN model for the first time. Theresults show that our work about real-time rapid prediction is successful.(3) Based on the ANN technique, the polar motion (PM) are predicted in twoways, i.e., using PM only and using both PM and AAM data. The results show thatfor 1 to 7 days forward prediction, the accuracy is not improved after introducingthe AAM data, but for 8 to 40 days forward prediction the accuracy is significantlyimproved after introducing the AAM data. Because the ocean and the ground waterare also the polar motion excitation sources except for the atmosphere, it awaitsfurther investigations for incorporating the atmosphere, oceans and ground waterinto the polar motion prediction.(4) Since June of 2006, the sea surface temperature anomaly (SSTA) of the westPacific has exceeded 0.5℃for 4 continuous months, it maybe evolve into a newENSO event. We apply the up-t0-date SSTA data to predict this event by AAN. Wepredict the SSTA of Nino_3.4 sea area, based on the monthly data released by theClimate Prediction Center (CPC) of NECP. We conducted 8 predictions fromSeptember, 2006 to April, 2007. Every prediction was made 12 months forward.We compare our results with those of CPC. Our work is closer to the average of alldynamical and statistical models. This demonstrates that our prediction has certainreliability and reference values.