A GNSS PWV filling and short-term forecasting framework fused hybrid neural network

Abstract

The long time series of precipitable water vapor (PWV) derived from the global navigation satellite system (GNSS) provides valuable information about atmospheric water vapor. However, existing long time series of PWV exhibits considerable data missing, and the short-term forecasting of PWV is insufficiently investigated, which becomes the focus of this paper. Accordingly, a hybrid driving framework is developed based on physical constraints and neural networks (PWV-FSFnet) utilizing GNSS PWV and meteorological (MET) data. This framework is used for long-time-series PWV filling and short-term PWV forecasting. In this framework, the long-time-series PWV filling model is first proposed by combining the linear and nonlinear variations of PWV and the relationship between PWV and MET parameters. Moreover, a short-term forecast model of PWV is developed by combining convolutional neural network and long short-term memory to predict the PWV of the next 1–6H, which considers the spatio-temporal relationship between PWV and multiple MET parameters. The experiment is performed in Mainland China using 957 GNSS stations, 1614 MET stations, and 87 radiosonde stations over the period of 2017–2024. Statistical results show that the PWV-FSFnet framework enables high-quality filling of long-time-series PWV with average RMS of 1.45 and 2.52 mm for internal and external accuracy, respectively. In addition, PWV-FSFnet demonstrates strong robustness in predicting PWV across different seasons, months, PWV levels, and climate regions, and the average RMS of hourly PWV forecasts is only 2.72 mm. The results demonstrate the feasibility and effectiveness of the proposed PWV-FSFnet framework in filling and forecasting PWV, highlighting its strong application potential in GNSS meteorology.