Empirical comparison of five deep learning architectures for GNSS time-series data forecasting

Purpose. This study aims to propose an effective deep learning model for analyzing time series of GNSS data. Methodology. Five deep learning models were investigated using the Adam optimization algorithm and the MSE loss function: CNN (Conv1D), GRU, Transformer, Seq2seq, and SimpleRNN. The models’ predictive performance was evaluated with data from three CORS stations in Vietnam: HYEN, BTRI, and CTHO. The models were constructed using the Adam optimization function, MSE loss function, and a batch size of 16. Findings. The prediction results across all models were generally low. Specifically, the Seq2seq model achieved the lowest RMSE value for the Up-component data from the HYEN station. The Transformer model was unsuitable for analyzing GNSS data over time, and the SimpleRNN model showed poor performance with the HYEN station data. However, the proposed CNN(Conv1D)-RF model, which integrates the RF model with the CNN(Conv1D) model, achieved high predictive performance, with RMSE = 0.4 mm, MAE= 0.07 mm, and R-squared = 99.8 % when the input data consisted solely of the Up-component. When the input data included all three components (N, E, and Up), the maximum RMSE value was 0.92 mm. Originality. This study introduces the CNN(Conv1D)-RF model, a novel integration of the RF model with the CNN(Conv1D) model, for GNSS data analysis. Practical value. The proposed model demonstrates promising predictive performance, indicating its potential application in accurately analyzing GNSS data for various practical purposes.