Deep learning-based modeling and prediction of GNSS time series: A comparative analysis of adaptive optimization algorithms

In this research, optimization algorithms with adaptive learning rates on Global Navigation Satellite System (GNSS) time series data are comparatively investigated. For this purpose, five years of GNSS measurement data obtained from the AGRD station located in the Ağrı province of Türkiye were used and incorrect or missing records were detected for a total of 251 days in the dataset. After the missing data were completed using the linear interpolation method, a total of ten different deep learning methods and four different adaptive optimization algorithms (Adam, Adagrad, RMSprop and AdamW) were used to develop separate prediction models and performance evaluations were performed. When the performance of the best combination, the Adam optimized-GRU model, was evaluated based on Root Mean Square Error (RMSE) values, it was found to be 1.58 mm, 1.36 mm and 3.07 mm for the north, east and up components, respectively. When evaluated according to the Mean Absolute Error (MAE) value, it was found to be 1.20 mm, 1.05 mm, 2.33 mm, respectively. As a result of the comprehensive analyses, it has been revealed that Adam and AdamW algorithms are more effective than the others among the adaptive optimization algorithms examined and the deep learning models optimized with these algorithms exhibit superior prediction performance on GNSS time series data. It is thought that the results obtained from this study will be an important reference on adaptive learning optimization algorithms for future studies in the field of GNSS time series and deep learning and will guide the research on the subject.