基于长短期神经网络的遥测参数趋势预测

时间序列分析的主要目的是根据已有的历史数据对未来进行预测。传统的时间序列预测主要依靠基于模型的方法,比如季节性差分整合移动平均自回归模型(SARIMA)和指数平滑法(EXP)等。此类方法的参数选择严重依赖于专家经验,适用性并不广泛。针对周期性遥测参数,采用长短期记忆网络(LSTM),学习长时序依赖关系并给出多步预测值。试验通过将预测问题转化为监督学习问题建立半实时仿真环境,并重点研究了观测窗口、预测窗口、网络结构等对性能指标的影响。对比LSTM、SARIMA、EXP,结果表明LSTM具备优异的线性拟合能力和良好的非线性关系映射能力。LSTM预测方法摆脱了传统方法受制于专家经验和模型精度低等问题,为开展实时遥测参数预测奠定了基础。

Trend Prediction of Telemetry Parameters Based on Long Short-Term Neural Networks

The main purpose of time series analysis is to predict the future based on the existing historical data.Traditional methods of predication of time series mainly rely on model-based methods,such as seasonal differential autoregressive integrated moving average model (SARIMA) and exponential smoothing method (EXP).The parameters selection of model-based method is limited by expert experience,and the applicability is not extensive.Aimed at cyclical telemetry parameters,the long short-term memory (LSTM) neural network is used to learn the long-term dependence of parameters,and to provide with multi-step prediction values.The semi-real-time simulation environment is established by transforming prediction into supervised learning,and the influences of observation window,prediction window,and network structure on performance indicators are studied.Comparing LSTM with SARIMA and EXP,the results show that LSTM has strong linear fitting and good nonlinear mapping abilities.The LSTM model used in cyclical telemetry parameter prediction gets rid of the problems of traditional methods being restricted by expert experience and low model accuracy,which provides the basis for real time telemetry parameter prediction.