Forecasting Satellite Attitude Volatility Using Support Vector Regression with Particle Swarm Optimization

摘要

—Forecasting the volatility of satellite attitude is a meaningful but complicated problem due to the complex non-linear characteristics of the standard deviation series, which reflects the volatility of satellite attitude. Support Vector Regression (SVR) is an efficient machine learning technique derived from statistical learning theory and has been successfully employed to solve regression problem of time series with nonlinearity in recent years. However, the generalization capacity of SVRs is greatly depend on their hyper-parameters and the process of tuning parameters manually is time-consuming. Particle Swarm Optimization (PSO) is a simple but effective optimization method inspired by social behavior of organisms such as bird flocking and fishing schooling. Thus, this paper proposes a hybrid PSO-SVR model to predict the volatility of these three attitude angles in satellites: Pitch Angle (PA), Roll Angle (RA) and Yaw Angle (YA), respectively. Thereinto, PSO is exploited to seek the optimal parameters for SVR to achieve satisfactory generalization capacity. The standard deviation series generated from telemetry data of Attitude Control System belonging to a Chinese satellite was used as experimental data to testify the performance of our proposed PSO-SVR model. The experimental results indicate that the hybrid PSO-SVR model can be a promising alternative to forecast the volatility of satellite attitude with relative high accuracy compared with grey model, residual grey model, and BP neural network.