A new tropospheric delay combination prediction model based on time series decomposition and deep learning

Abstract

Abstract Zenith tropospheric delay (ZTD) prediction is of great significance for high-precision navigation. However, ZTD modeling has proved to be challenging due to the presence of linear and nonlinear characteristics. In this paper, we propose a combination ZTD prediction model (SLA), which considers the trend-based and seasonal variations respectively. It decomposes ZTD time series via seasonal-trend decomposition procedure based on loess (STL), individually predicting nonlinear components with long short-term memory network (LSTM) and linear components with autoregressive integrated moving average model (ARIMA). Finally, the individual predictions are recombined. The SLA model is compared with LSTM, extreme learning machine model (ELM), ARIMA, and the empirical global pressure and temperature (GPT3) model. The SLA model shows the best result in all models by analyzing the evaluation indicators including root mean square error (RMSE, 1.32 cm), the average normalized root mean square error (NRMSE, 0.56%), mean absolute error (MAE, 0.98 cm) and the mean coefficient of determination (R2, 0.83). In addition, the data of different months was tested separately, and the result showed that the SLA model has the best performance of ZTD prediction. Moreover, the SLA model has good results up to 12h, with RMSE < 1.60 cm, NRMSE < 0.7%, MAE < 1.25 cm, and R2 > = 0.75. This study provides a new model to predict the ZTD, which is helpful for the precise positioning of GNSS and can be further applied in the study of meteorology.