Integrated signal decomposition and machine learning for GNSS height time series prediction: A performance evaluation

High-accuracy Global Navigation Satellite System (GNSS) height time series and predictions provide essential references for international terrestrial reference frame establish, crustal deformation monitoring, sea level change assessment and other geodynamic process. In this research, 11 machine learning models with 5 signal decomposition algorithms on GNSS time series prediction are comparatively investigated. For this purpose, we used 25-year GNSS height time series from 13 globally distributed IGS stations with less than 2 % data gaps in the dataset for prediction experiments. Firstly, Performance evaluation based on the evaluation error metrics was conducted on the selected 11 single-models, the Support Vector Machine (SVM) outperformance to the rest models, as an example of BGIS, it was found to be 2.45 mm, 0.05 mm, 3.12 mm, 0.74 and 1.23 for the MAE, MAPE, RMSE, R2 and WQE respectively. Secondly, by integrating five signal decomposition algorithms—Singular Spectrum Analysis (SSA), Ensemble Empirical Mode Decomposition (EEMD), Variational Mode Decomposition (VMD), Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Empirical Mode Decomposition (EMD)—with the SVM model (the optimal standalone model), this study demonstrates that signal decomposition substantially enhances the prediction accuracy of GNSS height time series. Experimental validation identifies the SVM–SSA model as the top performer for long-term prediction; for instance, at the BGIS station, it achieves a WQE of 0.11 and an R² of 0.99, markedly outperforming the standalone SVM model (WQE = 1.23, R² = 0.74). Finally, this study supplements experiments on hybrid models combining signal decomposition algorithms with (Convolutional Neural Network) CNN and Transformer models, and comparative analysis across 13 globally distributed IGS stations shows that the SVM–SSA model attains the optimal performance, with a mean R² of 0.98 and a mean WQE of 0.20, compared with other hybrid models (e.g., EEMD–CNN, WQE = 0.46; EEMD–Transformer, WQE = 1.28). These results collectively indicate that the SVM exhibits more stable learning capability and stronger generalization within the decomposition–prediction framework, rendering it suitable for high accuracy GNSS height forecasting across diverse global scenarios.