Reducing GPS orbital and clock errors using recurrent neural networks

Because satellite orbits and clocks are usually treated as known quantities in GPS positioning models, these errors are one of the most impactful in Global Navigation satellite system (GNSS) positioning, especially in single point positioning mode. In this work we tried to reduce orbit and clock errors using deep learning technique (recurrent neural network) by creating a model that predict the offsets of broadcast ephemeris and clock biases, from precise products. We tried two different methods, the first one is predicting the future offsets for the next +72 hours basing on both precise and broadcast ephemeris, and the second method is predicting the current offsets using only broadcast ephemeris as inputs. To create the models, we used Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures. These architectures are generally used to solve sequential problems in deep learning. We got results that vary between methods in terms of satellites and time steps. For the first method we get (10-50cm) and (0.5-1.7ns) reduced from (~130cm, ~14ns), It increases over time steps, and we get a minimum of (44cm) for the second method.