Research On The Spacecraft Time Series Prediction Method Based On LSTM

Spacecraft are sophisticated,expensive vehicles that are operated in complex environments.The establishment of predictive systems for spacecraft allows for better monitoring and control of spacecraft,avoiding risks and reducing losses.Telemetry data are collected by various sensors on the spacecraft and sent to the ground control center.It covers specific information on the operation of the spacecraft and can reflect the operational status of the spacecraft.Among the spacecraft telemetry data,there is a class of data with high volatility and non-stationary characteristics.It has dependencies and cooperativities with other telemetry data dimensions,and can be difficult to predict accurately using common single-dimensional telemetry prediction models.In this paper,a two-year multidimensional telemetry time series of a satellite is used as the research object to study the time series prediction of multidimensional spacecraft.The main work includes the following aspects:(1)An analysis of the research methods in the field of telemetry data prediction in recent years is conducted,summarizing the characteristics of different models.In this paper,the multidimensional telemetry data of spacecraft is analyzed,and its characteristics are summarized in terms of time dimension and related dependencies.At the same time,the telemetry data is preprocessed,and a high-quality telemetry data set is constructed,which provides the basis for subsequent predictions.(2)To address the prediction issues such as the high environmental impact of some telemetry data in spacecraft and its dependence on other telemetry dimensions,a realtime prediction model of multi-dimensional telemetry time series based on 1D Convolution,Bi LSTM and attention mechanism is proposed,with reference to mainstream multivariable time series method.In this paper,one-dimensional Convolution is used to capture the short-term dependence in time series;Bi LSTM is used to capture the bidirectional long-term dependence of time series data;attention mechanism in the dimension direction is introduced to screen the correlations between multidimensional telemetry data;then an AR model is added as a linear component to the prediction model,finally get the prediction of telemetry data.(3)Regarding multidimensional telemetry data prediction model,this paper uses two years of real spacecraft telemetry data for the experiment.A variety of evaluation indicators are selected to demonstrate the functionality and necessity of different components of the designed model through variant model comparison experiments.The prediction accuracy of the designed model is demonstrated by benchmark method comparison experiments.The experimental results show that the hybrid prediction model based on 1D Convolution,Bi LSTM and attention mechanism has significant performance advantages.And the model is a real-time prediction model that can be applied to actual aerospace prediction systems.