基于数据驱动方法的在轨卫星智能温度预测

Data-driven-based approach for intelligent temperature forecasting of in-orbit satellites

针对传统卫星温度预测方法在预测精度和鲁棒性方面表现不佳,难以满足高维度耦合数据预测需求的问题,提出一种针对卫星温度遥测数据的多元时序数据预测模型——改进的时间序列处理模块(advanced time series processing module,ATSPM)-Net。首先,构建了包含一维卷积和门控循环单元(gated recurrent unit,GRU)的ATSPM,以对高度耦合的遥测数据中的时间依赖关系进行多尺度提取。接着,设计了多元时序数据预测模型ATSPM-Net。通过堆叠ATSPM,ATSPM-Net确保模型的灵活感受野,从而实现高准确率和鲁棒性的遥测数据预测。最后,在5个数据集上进行的数值实验结果表明,相较于其他类型的时序数据预测模型,ATSPM-Net在参数量较少的情况下能展现出更优异的温度预测性能。

Aiming at the problem of poor prediction accuracy and robustness of traditional satellite temperature forecasting methods,which are difficult to meet the demand for high-dimensional coupled data forecasting,a multivariate time series data forecasting model for satellite temperature telemetry data—advanced time series processing module(ATSPM)-Net is proposed.Firstly,an ATSPM consisting of one-dimensional convolution and gated recurrent unit(GRU)is constructed to extract temporal dependencies from highly coupled telemetry data at multiple scales.Next,a multivariate temporal data forecasting model ATSPM-Net is designed.By stacking ATSPM,ATSPM-Net ensures the flexible receptive field of the model,thereby achieving high accuracy and robustness in telemetry data forecasting.Finally,numerical experiments conducted on five datasets showed that compared to other types of time series data forecasting models,ATSPM-Net can demonstrate better temperature forecasting performance with fewer parameters.