Accurate and reliable location information is essential for the safe operation of unmanned transport trains. However, in steel production and transportation environments, multipath interference and signal blockage often produce numerous pseudo-fixed solutions in GNSS-RTK positioning data, leading to unreliable train position estimates and posing serious safety risks. To tackle these challenges, this paper presents a robust GNSS-RTK localization method that integrates state estimation with trajectory prediction. Initially, the double-difference positioning model combined with the least squares algorithm is employed to obtain the GNSS-RTK float solution. This float solution is then fixed to achieve centimeter-level accuracy using an ambiguity decorrelation algorithm and ratio test. To overcome the limitations of the fixed threshold in the ratio test—which can result in pseudo-fixed solutions under multipath interference and signal obstruction—this approach diverges from conventional multi-sensor cross-validation methods. Instead, it leverages the spatiotemporal characteristics of the train′s motion by comparing the current position against the predicted trajectory derived from previous states. This enables rapid identification and correction of pseudo-fixed solutions without requiring additional hardware, thereby enhancing positioning reliability. The method was validated across semi-occluded environments, urban canyons, and steel transportation sites. Experimental results demonstrate its superior ability to detect pseudo-fixed solutions accurately while maintaining high-precision localization. Compared to threshold- and trajectory prediction-based anomaly detection methods, it achieves higher recognition rates, ensuring dependable train positioning in real-world industrial scenarios.