Comparative Evaluation of Regression Models and Bi-GRU for GNSS Position Time Series Forecasting

Abstract

Accurate forecasting of the up component in Global Navigation Satellite System (GNSS) time series is crucial for applications in geodesy, structural monitoring, and hazard mitigation. This study compares the performance of several regression models and a Bidirectional Gated Recurrent Unit (Bi-GRU), an architecture that shares similarities with Long Short-Term Memory (LSTM) networks but offers a streamlined structure, in predicting the Up component across ten GNSS stations located in Turkey and Europe. The models assessed include Bayesian Ridge Regression, Huber Regressor, Theil-Sen Regressor, Ridge Regression, Least Angle Regression (LARS), Automatic Relevance Determination (ARD) Regression, Gaussian Process Regressor, Light Gradient Boosting Machine (LightGBM) Regressor, and Random Forest Regressor. The results indicate that regression models like Bayesian Ridge, Huber Regressor, and ARD Regression generally provided more stable and accurate predictions than the Bi-GRU network, delivering reduced RMSE and MAE values alongside enhanced ${R^{2}}$ scores. This study demonstrates that simpler, regularized regression models can outperform complex neural networks such as GRU for GNSS time series forecasting, especially when dealing with noisy up component data.