关键词: GNSS, 大气可降水量, ERA5再分析数据集, LSTM模型

Abstract:

Three precipitation events in August 2022 in Harbin were selected as the research objects, and the atmospheric precipitable water (PWV) was calculated and retrieved based on the original GNSS observation data provided by the city's continuous operation reference station (CORS). The atmospheric precipitable water of ERA5 reanalysis dataset was calibrated using IGRA2 sounding dataset. The calibration ERA5 PWV was used to verify the inversion accuracy of GNSS PWV. The LSTM model was used to train PWV and analyze the temporal variability of PWV during the precipitation process to provide a reference for improving the monitoring ability of extreme precipitation events in Harbin and formulating countermeasures before disaster. The results showed that: (1) based on the calibration data of Harbin sounding station, the correlation coefficient of IGRA2 PWV and ERA5 PWV was 0.99, the mean deviation of IGRA2 PWV and ERA5 PWV was −0.91mm, the standard deviation was 2.01mm, and the RMSE was 2.20mm. It showed that ERA5 reanalysis data set had higher accuracy in Harbin. (2) The difference between GNSS PWV and ERA5 PWV was generally between ±4.00mm, the maximum mean deviation was −0.19mm, the maximum standard deviation was 1.51mm, and the maximum root mean square error was 1.52mm, indicating that GNSS had high accuracy in PWV inversion. (3) Based on the PWV value of GNSS inversion as the training data source, the RMSE of the training set of the constructed LSTM model was 0.73mm, indicating that GNSS PWV could train the LSTM model effectively. (4) The difference between LSTM PWV and GNSS PWV was generally maintained at ±5.00mm 12h before the precipitation event, and increased from 12h to 24h with the forecast step, and the difference between the two gradually increased to ±20mm. The variation trend of PWV was predicted by LSTM model and conducting short−term predictions of precipitation events within 12 hours was feasible.

Key words: GNSS, Precipitable water vapor (PWV), ERA5, Long short?term memory (LSTM) model