Prediction of ZTD based on a parameter separation strategy and a graph convolutional network method for regions with sparse GPS stations

Abstract

ZTD (Zenith Tropospheric Delay) prediction plays an important role in GPS (Global Positioning System) positioning and weather monitoring. However, accurate ZTD prediction remains a major challenge due to the differences in the characteristics of its components, and its application is also limited in regions with sparse GPS stations. ZTD mainly consists of ZHD (Zenith Hydrostatic Delay) and ZWD (Zenith Wet Delay), whose fluctuation characteristics differ significantly from each other. Therefore, a parameter separation strategy is proposed to predict these two components separately by combining a period component extraction method with the Informer model to improve prediction accuracy and efficiency. Moreover, to obtain ZTD values at any location in regions with sparsely distributed GPS stations, a Graph Convolutional Network (GCN) method is adopted in this study. The performance of the proposed method is evaluated using datasets collected from 2013 to 2023 at eight GPS stations in Hong Kong, China. Experimental results show that: (1) after applying the parameter separation strategy, the MAE (Mean Absolute Error) and RMSE (Root Mean Square Error) of the predicted ZTD decreased from 8.77 mm to 7.61 mm and from 12.69 mm to 6.68 mm compared to the traditional method. The improvement rate is approximately 13.23% for MAE and 47.36% for RMSE, respectively. (2) The ZTD prediction accuracy at any point using the GCN method achieves about 8.80 mm for MAE and 12.88 mm for RMSE, which is comparable to that of the traditional method based directly on GPS observations. (3) The prediction accuracy of ZTD is strongly affected by seasonal variations. The MAE and RMSE are larger in the rainy season than in the non-rainy season. Specifically, the highest prediction accuracy occurs in winter, moderate accuracy in spring, and the lowest accuracy in summer and autumn. In conclusion, the proposed method effectively improves ZTD prediction accuracy and can be applied in regions with sparsely distributed GPS stations.