Quasi-real-time retrieval of ERA5 precipitable water vapor over mainland China

The occurrence of extreme weather events is closely related to the spatial distribution and temporal variation of atmospheric water vapor. High-precision and high-spatiotemporal-resolution precipitable water vapor (PWV) data can effectively capture such variations. However, current approaches for obtaining PWV data still have certain limitations: although PWV retrieved from Global Navigation Satellite System (GNSS) observations enables all-weather real-time monitoring, its sparse ground station distribution hinders the achievement of high spatial coverage; satellite remote sensing, while offering wide-area spatial coverage, suffers from low temporal resolution and is susceptible to cloud cover and meteorological events, limiting its ability to provide temporally continuous and spatially complete PWV fields. The fifth-generation global atmospheric reanalysis dataset (ERA5) from the European Centre for Medium-Range Weather Forecasts (ECMWF) offers high spatiotemporal resolution and has demonstrated significant potential in meteorological applications. However, it has a data latency of approximately 120 h. To address the real-time demand for water vapor data in severe weather forecasting, it is essential to develop predictive models for ERA5 water vapor. Taking mainland China as a case study, given its vast geographic span and significant climatic and geomorphological heterogeneity, the study area was divided into 13 regions based on climate types, latitude, and landform characteristics. The Fast Fourier Transform (FFT) was employed to extract the common period of the ERA5 meteorological elements. The best common period for each meteorological element was identified through correlation analysis, and a temporal sliding window, corresponding to the best common period, was constructed to enhance the representation of spatiotemporal heterogeneity among the elements. To address the temporal delay in ERA5 data acquisition, the Convolutional Long Short-Term Memory (ConvLSTM) network was employed to predict ERA5 PWV across different seasons. Results show that among the 13 subregions of mainland China, a medium-length common period (e.g., 83 h) yields the best predictive performance. Topographic and climatic characteristics have a significant impact on prediction accuracy: the plateau region demonstrates better predictive performance due to stable water vapor and low atmospheric pressure, whereas the tropical monsoon region exhibits strong variability driven by monsoon activity, with annual RMSE trends closely aligned with the seasonal variation of water vapor. The model was externally validated against GNSS-retrieved PWV (1 Mar 2020–28 Feb 2021). Driven by eight ERA5 variables (PWV, temperature, pressure and wind-related elements), it achieved Root Mean Square Error (RMSE)s of 2.83–8.08 mm—the minimum over the low-latitude first-step plateau mountains (LFA, plateau-monsoon) and the maximum over the mid-latitude second-step mountains (MSM, temperate-monsoon). The Correlation Coefficient (R) ranged from 0.82 (MSM-mountain) to 0.97 over the low-latitude second-step hills (LST, tropical-monsoon), demonstrating robust performance across contrasting terrains and climates. In addition, external validation using radiosonde-derived PWV (RS PWV) at four representative stations yielded comparable results, with RMSEs ranging from 2.69 mm (LFA) to 6.54 mm (MSM) and R values ranging from 0.94 to 0.97. The integration of multiple meteorological elements significantly improves predictive accuracy: the RMSE is highest when only water vapor features are used (up to 13 mm in multi-step winter forecasts); with the introduction of temperature, wind fields, and other features, predictive performance progressively improves, and the average RMSE reduction reaches 40.86% when all elements are included. The ERA5 water vapor prediction model proposed in this study enables quasi-real-time estimation of ERA5 water vapor with high accuracy and fine spatiotemporal resolution. This model provides valuable support for short-term severe weather forecasting.