Seasonal Forecasts for the Horn of Africa
Principal Investigator:
Paolo Mori
Affiliation:
Institute of Physics and Meteorology, University of Hohenheim
Local Project ID:
WRFSFHOA
HPC Platform used:
Hazel Hen of HLRS
Date published:
Numerical models of the Earth (atmosphere, land surface, oceans, criosphere) are necessary to produce accurate weather and climate predictions. The computational costs limit the resolution and the level of detail the that can be implemented in global models. However, limited-area models can be run at higher resolutions to improve the representation of reality, although only for a limited region. The main advantage of high resolution is a more realistic representation of topographic features as mountains, lakes, coastlines (Schwitalla et al. 2008; Prein et al. 2015). In addition, some phenomena don’t need parametrization once a threshold is reached, hence the related model uncertainties is reduced. At a scale below 4 km the convection parametrization can be turned off, as the micro-physics scheme is sufficient to handle precipitation properly: this is called the convection-permitting scale.
Due to the complex topography and the influence of different factors, summer rainfall in the Horn of Africa is difficult to forecast correctly (Nikulin et al. 2018; Korecha and Sorteberg 2013). As they are crucial for human activities in the region, improvements in their prediction are sought. Dynamical downscaling of state-of-the-art seasonal forecasts to the convection-permitting scale has not been performed yet, over Africa or elsewhere, due to the costs of an ensemble of convection-permitting seasonal simulations and the additional hind-casts necessary to calibrate the model output. This experiment is limited to a one-season model run: its purpose is to investigate whether some added value is provided by the raw model output (without calibration). Additionally, the characteristics of the ensemble that can be exploited to reduce the number of ensemble members or to improve a probabilistic forecasts are investigated.
downscale to 3 km resolution the SEAS5 seasonal forecasts for summer 2018
Study the influence of lateral boundary conditions and physics schemes on model behavior
Calculate the added value of the high-resolution model for precipitation
Ensemble set up: the ensemble was designed to address uncertainties caused by lateral boundary conditions (LBC), being able to compare them to model physics’ variability. The most important atmospheric physics schemes for precipitation-related processes, once the cumulus scheme is turned off, are: short-wave radiation scheme, micro-physics, planetary boundary layer.
Total: 4 LBC *4 PHYS = 16 ensemble members
Clustering of Ethiopia: the k-means clustering method has been applied to rainfall time series in order to identify regions with homogeneous climate and precipitation distribution (Figure 2)
Temperature: mean summer 2-meter temperature of SEAS5 four-member ensemble forecast and 16-member WRF ensemble are compared with ECMWF ERA5 reanalysis in the following table:
| ERA5 (°C) | SEAS5 mean (°C) | WRF mean (°C) |
Cluster 1 | 26.1 | 26.8 ± 0.2 | 27.1 ± 0.3 |
Cluster 2 | 22.1 | 22.1 ± 0.2 | 23.0 ± 0.2 |
Cluster 3 | 26.3 | 26.4 ± 0.1 | 27.7 ± 0.3 |
Cluster 4 | 29.5 | 29.0 ± 0.2 | 30.2 ± 0.3 |
Illustration 3: Total precipitation in summer 2018, (mm) regridded to the GPM grid.
a) GPM-IMERG (Huffman, 2017), b) Chirps (Funk et al., 2015), c) SEAS5 ensemble mean, d) WRF ensemble mean, e) Bias: SEAS5 mean – GPM, f) Bias: WRF mean – GPM
Precipitation: the picture 3 shows total precipitation values for summer 2018 (June July and August) in the Horn of Africa. It is apparent how the high-resolution affects the precipitation distribution over the Ethiopian highlands, with small scale variability clearly visible, opposite to the smoother field of the low resolution global model. However, the high-resolution model is not able to reduce the biases revealed by the global model: the WRF ensemble mean has very high total values in the vicinity and to the west of the highest peak in the region. The wet bias is larger over the southern part of the highlands than in the northern half, similar to the parent model, but the WRF accumulated values are significantly higher. A dry bias is present in the north-western part of the domain for both SEAS5 and WRF, but in the latter case a bigger area at the foot of the Ethiopian highlands is affected: time series reveal that the rain belt movement to the area is delayed by about 15-20 days in the downscaled run. The consequent dry and warm biases are still visible at the seasonal scale.
| GPM (mm) | SEAS5 mean (mm) | WRF mean (mm) |
Cluster 1 | 31.6 ± 0.8 | 40 ± 12 | 60 ± 20 |
Cluster 2 | 402 ± 4 | 420 ± 40 | 680 ± 130 |
Cluster 3 | 844 ± 5 | 950 ± 60 | 1000 ± 120 |
Cluster 4 | 475 ± 3 | 500 ± 40 | 530 ± 70 |
Sensitivity of model precipitation The lateral boundary conditions are usually not considered in RCM studies because one single GCM or reanalysis products is used as forcing. In this setting, the ensemble forecast allows to evaluate the different contributions to the ensemble spread (that is, the model variability) of physics parametrization and boundary conditions. The spatial distribution of the relative contribution to the ensemble spread is shown in Figure 4. A pattern is apparent: LBCs variability is larger than PHYS, being dominant over the highlands in Ethiopia, Kenya, Uganda, the Arabic peninsula and the Indian Ocean.
Illustration 4: Spatial distribution of spread_LBC – spread_PHYS normalized by total ensemble spread at grid-point level for accumulated (JJA precipitation. Positive values (red) indicate where LBC spread is at least 50% larger than PHYS spread, negative values (blue) the opposite.
Copyright: Institute of Physics and Meteorology, University of HohenheimDistribution Added Value: the distribution skill score is a tool to measure the likeness of two probability distribution functions (PDF). In this case model and satellite observations are compared. The PDF skill score is based on the distributions’ overlap. The distribution added value DAV (Soares and Cardoso, 2018) is the normalized difference between the high-resolution model’s skill Shr and the driving model’s skill Slr, normalized by Slr itself: DAV = (Shr - Slr) / Slr
A positive value indicates an improvement of the coarse model by means of the downscaling, whereas a negative DAV suggests a performance degradation. The possible range is from negative infinity to positive infinity.
Downscaling the SEAS5 model did not always improve the precipitation distribution, as the negative values in the DAV column reveal. Only in the cluster 1 there is a minimal, although different from zero, added value. All other clusters see a rather consistent degradation.
| DAV | DAV95 |
Cluster 1 | 0.09 ± 0.06 | 6.6 ± 5.4 |
Cluster 2 | -0.11 ± 0.05 | 3.8 ± 1.6 |
Cluster 3 | -0.18 ± 0.02 | 1.1 ± 0.9 |
Cluster 4 | -0.06 ± 0.03 | 1.3 ± 0.5 |
However, when the 95th percentile is considered (that is, the most intense events – DAV95 column), the WRF downscaling provides added value in all clusters. The large uncertainty is consequence of the occasional very small values of Slr, producing large fluctuations in DAV95 values, which are nevertheless always positive.
Picture 5 shows an example of the different precipitation distribution characteristics. The red curve shows the global model distribution: the amount of low-intensity precipitation events is rather overestimated, while heavy rainfall events are almost missing. The regional model (in blue), on the contrary, tends to overestimate the contribution of intense events. Finding a balance between the two opposite behavior helps to provide better forecasts, for example the likelihood of extreme events in the incoming months could be predicted more accurately.
Illustration 5: Example of probability distribution functions of daily precipitation in Cluster 2 for summer 2018: comparison of measurements and model output (single ensemble members). On the x-axis are precipitation amounts in mm, the y-axis displays the share the total precipitation that is provided during days with the given precipitation amount or, in other words, how many days with that much precipitation have happened. Measurements are the yellow curve, as a reference.
Copyright: Institute of Physics and Meteorology, University of HohenheimComputation: all simulations were performed using the Weather Research and Forecasting (WRF) model. Each model run used 100 nodes on the HLRS Hazel-Hen system, with 2 OpenMP thread and 12 MPI processes per node.
About 3 million core hours were dedicated to the spin-up tests and the remainder for the downscaling of the ensemble forecast. About 17 TB of partially post processed output files have been produced, 6 by the spin-up runs and 11 by the downscaling of the seasonal forecast.
The lessons learned can be summarized as follows:
The limited-area model follows closely the large scale circulation, including the longitudinal movement of the rain belt associated to the West-African monsoon.
The high-resolution ensemble provides a larger spread for precipitation, which is often enough to fix the under-dispersion in the global ensemble. The same does not apply to 2-meter temperature.
Variation of lateral boundary conditions produces a much larger precipitation variability than physics schemes in large areas of the domain. Any further study on seasonal forecasts using a regional climate model will have to take this into account in order not to disregard one of the largest sources of uncertainty.
Even though the current model set-up does not improve the model bias and the overall precipitation distribution is slightly degraded, there is measurable added value when predicting extreme events.
This study suggests that the most efficient way to set-up a limited-area ensemble (in particular for operational use) would be to find the most suitable physical schemes considering the region, the global and the regional models in use, varying only the boundary conditions to generate the ensemble.
Funk, Chris, Pete Peterson, Martin Landsfeld, Diego Pedreros, James Verdin, Shraddhanand Shukla, Gregory Husak, et al. 2015. “The Climate Hazards Infrared Precipitation with Stations--a New Environmental Record for Monitoring Extremes.” Scientific Data 2 (December): 150066.
Huffman, G., 2017:, GPM IMERG Final Precipitation L3 Half Hourly 0.1 degree x 0.1 degree V05, Greenbelt, MD, Goddard Earth Sciences Data and Information Services Center (GES DISC), Accessed: data download v06b, 10.5067/GPM/IMERG/3B-HH/05
Korecha, Diriba, and Asgeir Sorteberg. 2013. “Validation of Operational Seasonal Rainfall Forecast in Ethiopia.” Water Resources Research. doi.org/10.1002/2013wr013760.
Nikulin, Grigory, Shakeel Asharaf, María Eugenia Magariño, Sandro Calmanti, Rita M. Cardoso, Jonas Bhend, Jesús Fernández, et al. 2018. “Dynamical and Statistical Downscaling of a Global Seasonal Hindcast in Eastern Africa.” Climate Services. doi.org/10.1016/j.cliser.2017.11.003.
Prein, Andreas F., Wolfgang Langhans, Giorgia Fosser, Andrew Ferrone, Nikolina Ban, Klaus Goergen, Michael Keller, et al. 2015. “A Review on Regional Convection-Permitting Climate Modeling: Demonstrations, Prospects, and Challenges.” Reviews of Geophysics 53 (2): 323–61.
Soares, P. M. and Cardoso, R. M. (2018), A simple method to assess the added value using high‐resolution climate distributions: application to the EURO‐CORDEX daily precipitation. Int. J. Climatol, 38: 1484-1498. doi:10.1002/joc.5261
Schwitalla, Thomas, Hans-Stefan Bauer, Volker Wulfmeyer, and Günther Zängl. 2008. “Systematic Errors of QPF in Low-Mountain Regions as Revealed by MM5 Simulations.” Meteorologische Zeitschrift. doi.org/10.1127/0941-2948/2008/0338.
Skamarock, W. C., and Coauthors, 2008: A Description of the Advanced Research WRF Version 3. NCAR Technical Note NCAR/TN-475+STR, doi:10.5065/D68S4MVH.
Dr. Thomas Schwitalla, Dr. Kirsten Warrach-Sagi,Prof. Dr. Volker Wulfmeyer (all: Institute of Physics and Meteorology, University of Hohenheim)
M.Sc. Paolo Mori
Institute of Physics and Meteorology, University of Hohenheim
Garbenstrasse 30, D-70599 Stuttgart (Germany)
e-mail: Paolo.Mori [@] uni-hohenheim.de
HLRS Project ID: WRFSFHOA
October 2019