Considering ensemble spread improves rainfall forecast post‐processing

Abstract

Post‐processing is an essential step in improving rainfall forecasts from numerical weather prediction (NWP) models for hydrological prediction. While NWP models provide informative ensemble forecasts, this paper concentrates on unravelling the role of ensemble spread in rainfall forecast post‐processing. The ensemble link functions (ELFs) post‐processing method is developed by employing the log‐sinh transformation to normalise the skewed and censored rainfall data and linking the predictive distribution to the mean and spread of ensemble forecasts. We apply the ELFs and Bayesian joint probability modelling approach (BJP) to calibrate ensemble precipitation forecasts from the European Centre for Medium‐range Weather Forecasts (ECMWF) Integrated Forecasting System (IFS). We compare ELFs to BJP, which only takes the ensemble mean as an input. While both the ELFs and BJP can reduce bias and improve reliability, we show that ELFs tends to outperform the BJP in improving forecast skill in over 50% of the cases examined. It is found that the BJP tends to overestimate forecasts for extreme events, diminishing forecast skill and that the consideration of raw ensemble spread in the ELFs contributes to reliable forecasts. The most substantial skill improvement of the ELFs over BJP is observed for a moderate level of underlying skill in the raw forecasts, typically corresponding to lead times from three to seven days. Overall, ELFs effectively use information from both the ensemble mean and spread to calibrate NWP rainfall forecasts more effectively than can be achieved when ensemble spread is ignored.