SVR and ARIMA models as machine learning solutions for solving the latency problem in real‑time clock corrections

Abstract

Real-time precise point positioning (PPP) has become a prevalent technique in global navigation satellite systems (GNSS). However, GNSS real-time users must receive space state representation (SSR) products to correct for satellite clock, orbit, and phase biases. The International GNSS Service (IGS) provides GNSS users with real-time services (RTSs) through different real-time correction SSR products. These products arrive at the GNSS users with some latency, which affects the quality of real-time PPP positioning. The autoregressive integrated moving average (ARIMA) and support vector regression (SVR) models are used in this research to predict those corrections to eliminate the latency effect. ARIMA model reduces the standard deviation by 28% and 13% for GPS and GLONASS constellations, respectively, compared to the real-time solution, which includes the latency effect, the research simulated the latency effect and named it a forced-latency solution, and the SVR model reduces the standard deviation by 28% and 23% for GPS and GLONASS constellations, respectively. The results for the permanent GNSS stations used in this study across different years 2013, 2014, 2015, 2019, and 2021 show a mean reduction in the 3D positioning standard deviation by 13% compared with the forced-latency solution for the ARIMA solution and 9% for the SVR solution. The potential of both models to overcome the latency effect is apparent based on the findings.