The Regional zenith tropospheric delay correction and precipitable water vapor prediction models based on the GNSS, ERA5 and machine learning

Zenith Tropospheric Delay (ZTD), one of the critical parameters in Global Navigation Satellite System (GNSS) meteorology, holds significant value for climate research and early warning of meteorological hazards. However, GNSS-ZTD time series are often unavailable at certain GNSS stations due to data loss caused by equipment malfunctions and network disruptions. Although ZTD time series can be estimated at any time and location based on prediction models and reanalysis datasets, which are generally limited in accuracy. The GNSS-ZTD data during 2020—2022 from 18 GNSS stations in Hong Kong, along with the fifth-generation reanalysis dataset from the European Centre for Medium-Range Weather Forecasts (ECMWF Reanalysis v5, ERA5) were used in this study. The correction models of ERA5-ZTD were developed using multiple linear regression, back propagation neural networks (BPNN), random forest (RF), and extreme gradient boosting (XGBoost) to interpolate missing historical GNSS-ZTD. Furthermore, GNSS-based precipitable water vapor (PWV) prediction models were constructed. Results show that the root mean square errors (RMSE) of corrected ERA5-ZTD using the single-factor, three-factor, BPNN, RF and XGBoost are 11.47 mm, 11.42 mm, 10.67 mm, 6.34 mm and 5.19 mm, respectively, where XGBoost is the best. When the prediction time of PWV are 1 h, 3 h and 5 h, respectively, the BPNN outperforms XGBoost and RF, the average RMSEs of prediction PWV are 1l.55 mm, 2.34 mm and 3.04 mm, and their average Bias are -0.13 mm, -0.13 mm and -0.16 mm, respectively.