Post-processing East African precipitation forecasts using a generative machine learning model

Abstract

Existing weather models are known to have poor skill at forecasting rainfall over East Africa, where there are regular threats of drought and floods presenting significant risks to people's lives and livelihoods. Improved precipitation forecasts could help mitigate the effects of these extreme weather events, through enabling advanced preparation and allocation of resources where needed. Building on work that successfully applied a state-of-the-art machine learning method (a conditional Generative Adversarial Network, cGAN) to postprocess precipitation forecasts in the UK, we present a novel way to improve precipitation forecasts in East Africa. We address the challenge of realistically representing tropical convective rainfall in this region, which is poorly simulated in conventional forecast models. We use a cGAN to postprocess the European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecast System (IFS) forecasts, at 0.1 degree resolution and 6-18h lead times. We investigate how well this method can correct bias, create samples of rainfall with realistic spatial structure and produce reliable probability distributions. The cGAN predictions and original forecast are also further postprocessed using a novel neighbourhood version of quantile mapping, in order to leverage the strengths of conventional postprocessing methods and provide a strong baseline to compare against. Our results indicate that the cGAN significantly improves the diurnal cycle of the IFS, and improves on metrics such as mean bias, and Fractions Skill Score up to high (99.9th percentile) rainfall values. However the quantile-mapped IFS achieves slightly higher Equitable threat Score values at the grid scale. The cGAN ensemble exhibits a mixture of behaviours at different rainfall scales, with under-dispersion at low rainfall values and over-dispersion at high rainfall values. Overall our results demonstrate how the strengths of machine learning and conventional postprocessing methods can be combined, and illuminate where the benefits of machine learning for postprocessing lie.

Date of Award	19 Mar 2024
Original language	English
Awarding Institution	University of Bristol
Supervisor	Peter A G Watson (Supervisor) & Laurence Aitchison (Supervisor)