AI-Driven DGPS Correction Based on Residual Forecasting via Kalman Smoothing

Suktae Kang*, Sung Wook Yun, Wonhee Jung, Woogeun Ahn

This paper proposes an AI-driven Differential Global Positioning System (DGPS) correction method based on residual prediction using a Kalman filter. Conventional DGPS correction methods rely on linear interpolation between multiple reference stations or on single-base station data, which becomes significantly less effective as the distance between the rover and reference station increases. When the baseline exceeds 100 km, spatial decorrelation of atmospheric errors and satellite orbit biases causes severe degradation in positioning accuracy. To address this limitation, we introduce a novel approach that leverages AI to predict residual errors at the rover site, even without nearby reference stations. The proposed framework first collects historical residual patterns from post-processed Global Navigation Satellite System (GNSS) data across multiple regions and varying ionospheric conditions. A machine learning model is then trained to estimate residual corrections based on pseudo range measurement. These predicted residuals are fused with conventional DGPS solutions using a Kalman filter to generate an enhanced trajectory. Performance was validated using publicly available RINEX data and simulated long- baseline scenarios. Simulation results demonstrate a noticeable improvement in position accuracy in particularly under sparse reference conditions. The proposed method shows sub-meter level corrections even beyond 150 km baselines. This hybrid approach enables the extension of DGPS applicability to previously unsupported areas and reduces dependency on dense reference station networks.

Keywords: GPS, DGPS, Kalman-filter, artificial intelligence