Accurate weather forecasting is an important guarantee for social development, safe operation of cities, people's lives, and defense against water and drought disasters. In recent years, despite the significant progress in numerical precipitation forecasting, the accuracy of the forecast is still affected by the uncertainty of the initial field, the limitations of the model structure and the limitations of the parameterization method, and the coarser spatial resolution also restricts its wide application in the field of meteorology and hydrology.
Therefore, in this study, a statistical post-processing study was carried out based on U-net on the multi-model super ensemble data CNE of CMA, ECMWF and NCEP, and an ensemble forecast precipitation dataset was developed for the downstream of the Yangtze River and the Lower Rivers for the period of 2021-2022. After U-net revision, the spatial resolution was refined from 0.5° to 0.1°, the deterministic and probabilistic accuracy indexes were comprehensively improved, and the forecast period was effectively extended, which provided refined data support for the model construction in the demonstration area of the project.
In this study, we use the global medium-term ensemble forecast precipitation data from three weather forecast centers, CMA, NCEP and ECMWF, in the TIGGE database, all with a spatial and temporal resolution of 0.5°/6h, ensemble memberships of 30, 30, and 50, and foresight periods of 360h, 360h, and 384h, in that order.
The forecast operation times of CMA and ECMWF are at Universal Time (UTC) of 00 and 12 o'clock, and NCEP at 00, 06, 12 and 18 o'clock, and the numerical model of all three is a combined perturbation scheme of stochastic physical tendency and stochastic kinetic energy compensation. The data can be downloaded at https://apps.ecmwf.int/datasets/data/tigge/levtype=sfc/type=cf/.
Mainly using python tools, the steps are:
(1) Aggregate all the ensemble members of CMA, NCEP, and ECMWF together by equal weight integration to form a multi-model super-ensemble forecast data CNE;
(2) Adopt the bilinear interpolation method to refine the 0.5° spatial resolution of CNE to 0.1°;
(3) Construct an ensemble forecast statistical post-processing model based on U-net deep neural network for the downscaled CNE;
Development of statistical post-processing datasets for data forecasting in the lower Yangtze River and Lower Rivers for 2021-2022 based on the constructed model.
After the U-net revision, the categorical indicator False Alarm Rate (FAR) is reduced from about 0.70 to about 0.55, and the Equitable Threat Score (ETS) is improved from about 0.15 to about 0.3, and the average gain of the foresight period reaches 0.17 and 0.14, respectively. the quantitative indicator KGE' improves under most of the foresight periods, and the effective maintains the spatial structure of the data. The comprehensive accuracy index CAS is effectively improved, and the foresight period is effectively extended up to 54 h. Therefore, the present data can provide a high-quality forecast precipitation dataset for the lower reaches of the Yangtze River and the Lower Lower River region.
| # | number | name | type |
| 1 | 2021YFC3000100 | Lower Yangtze River Flood Disaster Integration and Control and Emergency De-risking Technology and Equipment | National key R & D plan |
This work is licensed under
CC BY 4.0 (Creative Commons Attribution 4.0 International License).
CSTR:11738.11.NCDC.NHRI.DB6789.2025
10.12072/ncdc.nhri.db6789.2025
