Research On Prediction Methods For Telemetry Time Series Of The Satellite

Satellite is a complicated and muti-functional system based on several sophisticated subjects, which has great impact on the development of our nation. As satellite failures are always accompanying abnormal parameter values, analyzing the inherent regularity and future trend of historical data can prevent latent satellite failures in advance. Time series prediction is an efficient approach to study the behavior of key parameters. Based on the study of several existing time series prediction model and the feature of satellite time series, we research the prediction model sutiable for the satellite data. In this thesis, the main contents are:Firstly, we conduct intensive research on several widely used time series prediction methods, including Neural Network(NN), Support Vector Regression machine(SVR), Revelence Vector Machine(RVM) and Grey Model(GM). Each has its own merits and disadvantages. NNs can handle nonlinear mapping problems but they are always stuck in local extremum, overfitting, and slow convergence. SVRs can effective deal with high-dimensional and nonlinear problem with small sample, but their free parameters are difficult to set. RVMs can give probabilistic output, but they confront the same problem with SVRs. GM can only predict time series with specific feature.Secondly, satellite time series own unique characteristics, such as huge quantity, data-intensive, slow-varying and existing outliers. We propose a rounded processing procedures including cleaning outliers, data compression, data transformation, and data normalization, according to suggestions from satellite monitoring personnel and our research into the satellite data.Thirdly, we propose a satellite attitude short-term prediction method based on the hybrid PSO-SVR model which utilizes the desirable searching capacity of PSO and excellent nonlinear mapping performance of SVR. PSO can be used to seek proper parameters(penalty coefficient, insensitive parameter, and kernel parameter) for SVR. Comparision experiments with BPNN, GM(1, 1) and residual modified GM(1, 1) conducted on real dataset from a satellite’s attitude control system indicate that our proposed method has higher prediction accuracy.Fourthly, RVM is a new learning algorithm based on Bayesian framework, which can obtain sparsity and probabilistic solution. We are inspired by the sporadic study on PSO-RVM model applied in other fields and present a new interval prediction method based on PSO-RVM for the satellite power system. As the same, PSO is used to seek proper parameter(kernel paramete) for RVM. IN1 and VN1 from a satellite power system are used as experimental objects. Compared with PSO-SVR, PSO-RVM model can achieve slight better prediction accuracy, more sparse solution and shorter test-time. Besides, our proposed method can obtain desirable prediction intervals.In summary, the study in our thesis can provide some assistance for satellite monitoring personnel and give a reference to other researchers interested in the satellite time series prediction.