Does a multivariate approach enhance univariate grid-to-point post-processed rainfall forecasts? A comparative analysis
Title | Does a multivariate approach enhance univariate grid-to-point post-processed rainfall forecasts? A comparative analysis
|
Technical memorandum
|
|
Date Published |
01/2025
|
Secondary Title |
ECMWF Technical Memoranda
|
Number |
924
|
Author | |
Publisher |
ECMWF
|
Abstract | Reliable and accurate point-scale rainfall forecasts are crucial for mitigating the impacts of localised extreme hydro-meteorological hazards. Nevertheless, biases and the low spatial resolution of global numerical weather prediction (NWP) models can hinder their utility for predictions at a local scale. Statistical post-processing is widely used to correct biases in raw global rainfall forecasts and can address representativeness errors. The most advanced post-processing techniques follow a multivariate approach, using relationships between rainfall observations and multiple predictors to capture physical processes that can influence the corrections applied to the target variable (i.e., raw rainfall forecasts). However, multivariate post-processing is expensive to run and maintain. Post-processing techniques can also follow a univariate approach, adjusting raw rainfall forecasts only through relationships between past forecasts and observed rainfall. Univariate approaches are widely used due to their simplicity and low calibration and production costs. They might also be the only viable approach when the required information to define predictors is unavailable. This study compares the reliability and discrimination ability of rainfall forecasts post-processed with a multivariate and univariate approach to assess whether the former improves performance. While not explicitly evaluated in this study, improved performance is a prerequisite to justifying the multivariate approach’s higher maintenance, calibration, and production costs compared to its univariate counterpart. In this study, we considered the post-processing technique developed at the European Centre for Medium-range Weather Forecasts (ECMWF) to produce probabilistic rainfall prediction at point-scale called ”ecPoint”. One global forecasting system providing outputs in three different ways, two of which involved postprocessing, were compared: the raw ECMWF ensemble (ENS), the original and currently operational multivariate ecPoint (M-ecPoint), and an experimental univariate ecPoint (U-ecPoint). Compared to U-ecPoint and ENS, M-ecPoint provides better reliability and discrimination ability. It also provides more timely and accurate predictions of ”dry” and ”extremely wet” conditions. Moreover, compared to U-ecPoint, M-ecPoint maintains high detection of low-probability, high-impact extreme rainfall events while reducing false alarms, delivering much higher user confidence. |
URL | https://www.ecmwf.int/en/elibrary/81638-does-multivariate-approach-enhance-univariate-grid-point-post-processed-rainfall |
DOI | 10.21957/eb6c62b7af |
Download citation |