A precipitation forecast model with a neural network and improved GPT3 model for Japan

Accurate monitoring of atmospheric water vapor content is essential for the early warning of extreme weather events. As known, GNSS zenith troposphere delay (GNSS_ZTD) is an indispensable data source for retrieving precipitable water vapor (PWV). However, the newest GPT3 empirical model is not accurate enough to perform the ZTD (GPT3_ZTD) and PWV (GPT3_PWV) estimation in some regions, such as Japan. Thus, here, we introduce a radial basis function (RBF) neural network to establish ZTD forecast models based on the GPT3 model and use the predicted ZTD to retrieve PWV and adopt the retrieved PWV in forecasting precipitation. To thoroughly verify the accuracy of forecast results in 2021, we selected three external validation data: GNSS, radiosonde, and meteorological data. The GNSS_ZTD validation results show that the error compensation model of GPT3 based on RBF is superior to the GPT3 model and the model using a single RBF and back propagation (BP) neural network. The average RMSE of all GNSS stations is 50.7 mm, 53.7 mm, and 37.8 mm for GPT3_ZTD, RBF_BP_ZTD, and RBF_GPT3_ZTD, respectively. The GNSS_PWV and RO_PWV validation results show that the retrieved PWV with compensation of RBF_GPT3_ZTD is better than the uncompensated GPT3_ZTD, the average accuracy of RBF_GPT3_PWV of GNSS stations and radiosonde stations is improved by 40.4% and 25.8% against that of GPT3_PWV. For the precipitation forecast model results, the average forecast accuracy of all GNSS stations and radiosonde stations is 63.1% and 61.4%, according to the ERA5 precipitation. The average forecast accuracy is 66.3%, validated by meteorological precipitation records. The proposed model not only improves the GPT3 model but also forecasts the PWV value, which can improve the precipitation forecast in Japan, and is expected to expand to other regions.

Automated summary

Accurate monitoring of atmospheric water vapor content is crucial for early warning of extreme weather events. This study introduces a radial basis function neural network to improve the GPT3 model and utilize the predicted zenith troposphere delay (ZTD) to retrieve precipitable water vapor (PWV) for precipitation forecasting. The validation results show that the retrieved PWV with compensation of RBF_GPT3_ZTD is more accurate than the uncompensated GPT3_ZTD, with an average accuracy improvement of 40.4% and 25.8%. The precipitation forecast model achieves an average accuracy of 63.1% and 61.4% for GNSS stations and radiosonde stations, and 66.3% validated by meteorological precipitation records.