Employing artificial intelligence in Galileo orbital error prediction for real-time offline positioning

Abstract

Over the past years, the capabilities of Artificial Intelligence (AI) have been leveraged to handle complex problems facing Global Navigation Satellite Systems (GNSS) positioning. Satellite orbital errors represent one of these problems, particularly in real-time applications. In numerous scenarios when mobile networks are unavailable, the direct approach of using the International GNSS Service ultra-rapid orbits is not feasible. This commonly emerges when operating in rural and remote areas where the receiving device is offline for prolonged durations. This study provides a solution during these scenarios by employing AI to predict the orbital errors of Galileo satellites for real-time positioning. Different deep learning architectures are trained and tested using data over almost six years to assess the performance that each architecture can provide. This includes Deep Neural Networks (DNNs), Convolutional Neural Networks, and Long Short-Term Memory (LSTM) Recurrent Neural Networks (RNNs). The training process involves the broadcast and final orbits of Galileo satellites, the solar and lunar positions, the solar and geomagnetic indices, and the satellite block types. The results show that the developed DNN architecture can provide the best performance with a promising prediction accuracy. The mean absolute orbital errors can be reduced by an average of 72.0% and a maximum of 75.4% (for satellite E21). Additionally, testing the model in real-time positioning shows that the epoch-wise improvement of positional solutions can reach 0.86m (49.4%) with an average of 0.17m (10.7%). These results emphasize the substantial role of employing the developed DNN in offline real-time positioning.