Modified Deep Transformers for GNSS Time Series Prediction
Mostafa Kiani Shahvandi, Benedikt S. Soja · 2021
Highly accurate time series prediction in the field of geodesy is both important and a demanding task. For this problem we have investigated the potentiality of deep transformers, a deep learning approach. We have slightly modified the original network architecture and the optimization procedure of this model and thus have created a deep learning regression framework. We have applied the method for time series prediction of more than 18000 GNSS stations. We show that this approach performs better than traditional statistical methods by 21.5% in prediction accuracy. Furthermore, we show that it outperforms other machine learning algorithms by at least 2.7%. We demonstrate that millimeter accuracy can be expected for the prediction.