Research On Prediction Model Of The Satellite Clock Bias Based On Dynamic And Static Neural Network

With the increasing demand for high-precision real-time navigation and positioning in various fields,the accuracy of satellite clock bias data is one of the important factors which restrict its wide application.It has great significance for high-precision real-time positioning to establish a high-precision clock bias predicted model.Artificial neural networks have great advantages in non-linear data processing.However,the random initialization of the model parameters in the network will lead to unstable modeling results,causing the target equation to fall into a local minimum.By simulating the evolution of human thinking,the MEA algorithm utilizes the similartaxis and dissimilation operations to have a strong global optimization ability and a fast convergence speed,which can solve the initial parameter selection problem of the neural network well.Compared with the static neural network.the dynamic neural network with feedback and memory has good dynamics and can be better applied to the processing of time series.It also provides a new method for clock bias modeling.By considering the shallow static neural network with the dynamic neural network algorithms,and aiming at the problem that the easy overfitting of the traditional BP neural network,and the advantages of NAR neural network in time series modeling and prediction,this paper mainly introduces the clock bias predicted model commonly used in satellite navigation,proposing satellite clock bias predicted model based on the MEA-BP algorithm and the NAR algorithm.And short-term and medium-and short-term prediction experiments were carried out through IGS precision clock bias products.The main research content and work of the article are as follows:(1)The QP model,SA-QP model,GM model and ARM A model are introduced.This paper conducted experiments on common models using actual GPS data,and the advantages and limitations of each model were compared and analyzed based on the predicted results.(2)Before the clock bias modeling and prediction,original clock bias data made once difference to obtain the corresponding once difference sequences which is substituted into the model for training,and the final clock bias data is obtained by restoring the predicted value.Comparing with the use of original data for modeling and prediction,the accuracy and stability of neural network modeling and prediction are significantly improved by using the clock bias data after once difference processing.(3)To solve the problem of BP neural network applied in satellite clock bias modeling,some improvement measures are proposed:the optimal number of hidden layer nodes is obtained by numerical simulation.Optimizing the initial values of weights and thresholds through the MEA algorithm,and the MEA-BP model suitable for clock bias prediction is constructed.and the specific steps of using this model for the clock bias prediction are given.The experimental results show that:1)The MEA-BP algorithm can avoid the conventional BP algorithm from falling into the local optimum,and has higher accuracy in clock bias prediction.2)In the four periods short-term prediction,the prediction precision obtained by using the MEA-BP model is better than 0.36ns.0.38ns.0.62ns and 1.56ns,and can be improved by the minimum of 9.09%,18.08%.16.39%and 24.35%,respectively.When the forecast duration increases,the prediction error curve is more stable,and the precision and stability of the new model in the short-term and medium-and short-term prediction of clock bias are better than the four traditional models,which shows good prediction performance.3)This paper conduct research on data with different fitting time and sampling interval.The model prediction precision gradually increases as the amount of fitted data increases.When the data sampling interval is 30s,the highest prediction precision is obtained.(4)The NAR dynamic neural network algorithm is introduced and taken advantage of in time series processing,a NAR network structure based on satellite clock bias prediction is constructed,and short-term clock prediction is carried out based on this.The result shows:1)Except for the relatively low precision of the G24 satellite prediction(may be related to the significant fluctuations in its data),the precision of the remaining satellite prediction is controlled within 0.35 ns.In the four periods clock bias prediction,the precision is generally improved by more than 23%,and the highest can be improved by 86.49%,89.44%,93.81%and 95.19%.The fluctuation of the prediction error curve is small,and the prediction performance of the NAR algorithm is better than the four traditional models,which can effectively predict the satellite clock bias.And there are fewer iterations during training,which can greatly improve the forecast efficiency,and showed the new model is better in practicability and stability in the short-term prediction of clock bias.2)The effects of delay order,fitting data amount and sampling interval on prediction precision are discussed.Improper selection of delay order will seriously affect the precision of clock bias prediction,and the maximum change in precision is in the ns level;In the modeling using the data from the first half of the day,the precision of model prediction increased with the increase of the fitted data;Compared with the data of 5min and 15min sampling interval,the 30s sampling interval can obtain relatively high prediction precision.