Prediction Of Length Of Day Based On Artificial Neural Networks

The Earth rotation parameters (ERP) prediction is of great significance for the theoretical study and practical application of astronomy and geodesy. For the earth rotation parameters, the most difficult to forecast is the length of day (LOD), which contains complex nonlinear factors. In all the existing scientific studies, LOD contains linear and nonlinear prediction. The paper proposes to forecast the LOD using nonlinear neural network (Genetic Algorithm (GA) optimized BP Neural Network, Generalized Regression Neural Network (GRNN)), and the results are compared with the existing analysis.The main contents of the paper are:(1) Analysising the basic method and the shortcomings of the BP neural network in forecasting LOD, proposesing optimizing BP neural network's initial weights and thresholds by the use of genetic algorithm, Then GA-ANN is used to predict the LOD, and the results are compared with BP neural network prediction results.(2) To obtain the optimal weights and thresholds, loop calculations are needed in prediction. The lengthy astronomical data will cost more prediction time, which have difficulties in accessing real-time data. This paper applies a more simple and efficient GRNN to forecast the LOD. It is a local optimization algorithm, and does not fall into local minima. The results are analyzed and compared with those obtained by Schuh (2002) and EOP PCC (2010), it shows that using GRNN to predict LOD is feasible.(3) The input architecture in forecasting the LOD with traditional methods is in the manner of lead time i (i for the interval, taking 1,2,3... n). The ANN simulates the correlations among things. It can get more prior information when the relevance is greater. As a result, the output will be more accurate. However, when forecasting LOD with ANN, some important prior information will be missed as the correlations between data are weakened. In this paper, LOD is predicted by ANN with the data sampled in continuous way, and compared with the data sampled in lead time i. It shows that the method of the sampled data in lead time i has higher accuracy in ultra-short term prediction, and the method of the sampled data by continuous way has higher accuracy in short and medium term prediction.