Research And Application Of Based On Extreme Learning Machine For The Prediction Of Satellite Clock Error

Time is one of the most fundamental physical quantity in nature. In fact, it also determines the development and the application of the satellite navigation and position system. Precision measurement of satellite navigation and positioning system of distance and position, is the essence of the precision measurement of time. Affected by some of error factors in the process of satellite positioning, such as the ionosphere delay,tropospheric delay, multipath effect, satellite clock error, these are the time available to measure. Since the satellite clock error has an important influence on the satellite navigation and positioning, to obtain high accuracy real-time satellite clock error in the short term need to do a lot of research work, such as improving the atomic clock stability, improving or exploring the prediction model etc. Therefore, this paper will focus on the satellite clock error and forecast research.For the satellite clock error prediction, several prediction models are commonly used: quadratic polynomial model, grey system model, time series model and neural network model etc. With the development and research of neural network model, this paper will apply Extremely Learning Machine(ELM)---- a new feed-forward neural network model, for satellite clock error prediction. Taking the satellite clock error as the data source, it is discussed that the ELM forecast performance depend on different input quantity in the case of study; Aiming at the deficiency of its existence for ELM,this paper focus on the aspects of the hidden layer node number optimization, looking for random parameters given the best optimized ELM, and comparing the effects of different incentive function on ELM prediction.In this paper, applying the quadratic polynomial model, grey system model and RBF neural network for different satellite clock prediction, with the ELM and its improved model of satellite clock error prediction results were compared. Forecast experiments show that, improved extreme learning machine forecast output in a day higher than other models of accuracy. The variance prediction error of the best in a day can reach 0.38 ns, and it has great advantage in the satellite clock error of forecast comparing to other models in the short term forecast. Compared with the prediction models, this neural network model has the advantages of simple structure, fast learning speed, and good generalization performance.