Comparison of Combined and Individual Model in GNSS Data with Recurrent Neural Networks

Abstract

In this study, the effect of model management on performance is analysed by comparing the performance of station-based models and the single model trained with all station data using the Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) deep learning algorithms for the North, East and Vertical components of GNSS station data. For Scenario I, where separate models are used for each GNSS station, and Scenario II, where a single combined model is used with aggregated data, model performance is evaluated for the East, North and Vertical components using Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Coefficient of Determination (R-2). With the GRU algorithm, the average RMSE for the East component is 1.68 and 1.67 mm and the MAE is 1.24 and 1.27 mm for scenarios I and II respectively; for the North component the RMSE is 1.70 and 1.72 and the MAE is 1.32 and 1.33 mm; for the Vertical component the RMSE is 4.50 and 4.43 mm and the MAE is 3.58 and 3.50 mm. The results show that the single model approach can simplify model management and achieve comparable accuracy to separately trained models, especially in regions with more homogeneous data characteristics.