Hybrid GNSS time-series prediction method based on ensemble empirical mode decomposition with long short-term memory

Abstract

Abstract GNSS time series prediction plays a crucial role in monitoring crustal plate movements, dam deformation, and maintaining the global coordinate framework. To address the shortcomings of traditional GNSS time series prediction methods including insufficient feature selection, limited stability, and low predictive accuracy, this paper proposes a prediction model that combines the Ensemble Empirical Mode Decomposition (EEMD) with Long Short-Term Memory (LSTM) algorithm. The model utilizes the EEMD method to obtain reconstructed signals, which are then input as features into the LSTM model to predict the original time series. To validate the performance of the EEMD-LSTM combined model in GNSS time series prediction, experiments are conducted using time series data from 10 GNSS observation stations. The experimental results demonstrate that the EEMD model can accurately and effectively extract data features. Compared with the EMD-LSTM model, the EEMD-LSTM model achieves an average reduction of 28.32% in RMSE values, an average reduction of 28.52% in MAE values, and an average increase of 24.24% in R2 values. The EEMD-LSTM model exhibits higher predictive accuracy and a strong correlation with the original time series, thus demonstrating its capability to forecast target time series effectively. Therefore, this prediction model holds significant application value in GNSS time series prediction.

Keywords

• GNSS height time series
• Prediction model
• EEMD
• LSTM