Rainfall forecasts in daily use over East Africa improved by machine learning

TL;DR

The performance of the forecast system, cGAN, is tested, which is the only high-resolution (10 km) ensemble rainfall product that does real-time, probabilistic correction of global forecasts for East Africa.

Abstract

Ensemble forecasting has proven over the years to be a vital tool for predicting extreme or only partially predictable weather events. In particular life-threatening weather events. Many National Meteorological Services in East Africa do not have the computing resources to enable them to run their local area models in full ensemble mode over the full period of the 2 week medium range. As a result, weather users in these countries are not being given sufficient information about weather risk that is needed to make reliable decisions about taking preventative action. Consequently, society in many parts of the world is not as resilient to weather events as they could be. In this paper we test the performance of our forecast system, cGAN, which is the only high-resolution (10 km) ensemble rainfall product that does real-time, probabilistic correction of global forecasts for East Africa. Compared to existing state-of-the-art AI models, our system offers higher spatial resolution. It is cheap to train/run and requires no additional post-processing. It is run on laptops and can generate many thousands of ensemble members at little computational cost (compared with physical local area models). It is ideally suited to Meteorological Services with limited computational facilities.

Figures (10)

• Figure 1: Left: The East Africa region. The forecast domain spans 13.7 degrees south to 24.7 degrees north and 19.1 to 54.3 degrees east. Right: Example map of the rainfall categories listed in table \ref{['table:dataCategories']}. The data from all grid points in a category is combined to train one rainfall post-processing model, for a total of 12 models.
• Figure 2: Difference between the CPRS of the 24h rainfall accumulation forecasts and the CRPS of the IMERG climatological distribution at two example lead times, 30h (top) and 126h (bottom), averaged over one year for left: IFS, middle left: IFS with cGAN post-processing to 1000 ensemble members, middle right: IFS with IDR post-processing and right: GraphCast with IDR post-processing. Blue means that the model has a lower (better) CRPS. Red means that the IMERG climatological forecast has a lower CRPS.
• Figure 3: 6h and 24h rainfall accumulation CRPS values averaged over the East African domain and the one year test period. The top four plots (a,b,c,d) represent the four different 6 hour rainfall accumulation periods per day. Times are UTC and all forecasts are initialised at 00:00 UTC. The bottom plot (e) represents rainfall accumulation from 06:00 to 06:00. Each point represents the start of the accumulation period. QM stands for quantile mapping applied to each of 50 ensemble members. 1000 ensemble members are generated for the IFS+cGAN points. The blue line is the CRPS of the IMERG climatological distribution. Lower is better, but the domain average masks important issues, see figure \ref{['fig:cGAN1000']}.
• Figure 4: Difference between the one year mean CRPS of the 24h rainfall accumulation of 1000 ensemble members of the IFS+cGAN forecast trained on data from all forecast lead times, denoted IFS+cGAN-all and left: IFS+cGAN trained on data only from the lead time plotted denoted IFS+cGAN lead time and right: single ensemble member (deterministic) GraphCast with IDR post-processing to obtain a distribution. Blue means that the labelled model has a lower (better) CRPS. Red means that the IFS+cGAN-all forecast has a lower CRPS.
• Figure 5: Histograms of the rainfall in the test period averaged over 24 hours using a 30h to 54h lead time. (a) Left: Rainfall below 1 mm/h. (b) Right: Rainfall above 1 mm/h. The dashed lines indicate the distribution over the model training period. The solid lines indicate the distribution over the model test period. Note the different logarithmic axes. The lines labelled cGAN indicate cGAN applied to post-process the IFS and the brown line indicates a further correction by quantile mapping, IFS+cGAN+QM.