A guest blog by the Ad hoc Blue Nile Forecast Group (listed alphabetically): Sarah Alexander (1), Paul Block (1), Annalise Blum (2), Shraddhanand Shukla (3), Shu Wu (1), Temesgen Yimane (2), Ben Zaitchik (2)*, and Ying Zhang (2).
- University of Wisconsin-Madison, Madison, WI, USA
2. Johns Hopkins University, Baltimore, MD USA
3. University of California Santa Barbara, Santa Barbara, CA, USA
*Correspondence can be addressed to firstname.lastname@example.org
Ethiopia will soon begin filling the reservoir of the largest hydropower dam in Africa, the Grand Ethiopian Renaissance Dam (GERD). Construction of the GERD has been highly controversial because it sits on the Blue Nile River (Figure 1), which provides about 60% of total natural Nile flow into Egypt. Egypt relies almost exclusively on the Nile River for its renewable freshwater supply. When complete, the GERD will rise 155 meters tall, have an installed generating capacity of over 6,000 megawatts, and create a 74 cubic kilometer reservoir. This dam will be the first major infrastructure project in Ethiopia on the main stem of the Blue Nile.
Figure 1: The Nile Basin (yellow), including the GERD site (orange star) and GERD catchment (green)
The history, politics, and current diplomatic impasse regarding the construction and operation of the GERD have been covered extensively and sometimes breathlessly in media outlets around the world (for example: here, there and everywhere). We won’t get into those issues here, but suffice it to say that there is reason for concern. This is true both because of the actual hydrological impacts that filling of the GERD reservoir might have on downstream countries and because the dam has become a focal point for broader geopolitical tensions. In this context, we ask a simple forecast question: what is the status of seasonal forecasts of Blue Nile River flow? Few would disagree that filling the reservoir during wet years would lead to fewer acute impacts on downstream countries. While it is not clear whether Ethiopia will pursue a forecast-based filling policy (valuable as that might be!), we believe it would be useful for interested Parties to share a similar understanding of how the rainy season is likely to unfold in each year of the filling period if and when they engage in discussions of filling plans and responses.
With this as motivation, this post will: (1) offer a brief review of the drivers of rainfall variability in the Ethiopian Blue Nile basin; (2) describe existing statistical and dynamically-based forecast systems; and (3) offer an informal ensemble prediction for 2018 flows. The objective of this prediction is not to supplant existing forecast systems in the region. High quality meteorological seasonal forecasts are available from the Greater Horn of Africa Climate Outlook Forum (GHACOF) as well as from the Ethiopian National Meteorological Agency and other nations’ forecast centers. Rather, we aim to present a broad range of predictions generated using different methods, and generated in the absence of any perception of subjective input.
For those in a hurry: our multi-method seasonal forecast ensemble shows a strong likelihood of average to above-average Blue Nile basin rainfall and associated Blue Nile River flows at the GERD site in 2018 (Figure 2). The result holds true across statistical forecasts and dynamically-based forecasts (Figure 3). This is consistent with regional seasonal outlooks issued by several independent modeling centers. If the Ethiopian government does begin to fill the reservoir this year, then, these forecasts suggest that there is a low probability that reasonable filling activities would lead to catastrophically low flow in Nile River downstream. This seasonal outlook does carry substantial uncertainties, as we describe below.
Figure 2: Percentage of forecasts in this study predicting below normal, near normal, and above normal June-September rainfall (left) and June-December streamflow at the GERD site (right) for 2018, including eight NMME models (ensemble mean for each model) and eight statistical models. NMME forecasts are adjusted for mean biases. No variance adjustment was applied to any of the model forecasts.
Figure 3: (A) Boxplots of June-September historical rainfall, 1982-2017, according to CHIRPS rainfall estimates (grey), and of the 2018 forecasts for dynamically-based models participating in the North American Multimodel Ensemble (NMME; green) and for statistically-based models applied in this study (blue). Dashed lines show the upper and lower terciles of historical rainfall totals. (B) As in (A), but for the June-December forecast Blue Nile flow at the GERD site.
Drivers of rainfall variability in the Blue Nile basin
The Ethiopian Blue Nile (or Abay River, to Ethiopians) is located in the western portion of the country, in a region that gets the vast majority of its rainfall in the summertime kiremt season, between June and September. The flow of the Nile at the GERD site is highly seasonal, following the rains. Peak flows are most commonly seen in October (Figure 4A). Precipitation in the Blue Nile is variable from year to year (Figure 4B), and this rainfall variability is associated with strong interannual variability in Blue Nile flows (Figure 4C). The coefficient of variation for annual rainfall in the basin is on the order of 6%, while the coefficient of variation in annual flow near the GERD is close to 20%.
Figure 4: (A) climatological monthly average rainfall for the GERD catchment area (CenTrends data) and climatological monthly average Blue Nile River flow at El Diem, near the GERD site; (B) density function of annual total CenTrends rainfall for the GERD catchment; (C) as in (B), but for El Diem flows. For all plots rainfall data are for the period 1965-2009, constrained by our access to flow data
The effort to characterize, explain, and predict this variability dates back thousands of years, to the Nilometers of Ancient Egypt and the stories of the Bible. It has also been an active focus of modern climate research, and we won’t attempt a review here; see Berhane et al. (2013) or Nicholson (2017), among others, for more comprehensive treatment. Some prediction-relevant highlights from the literature on drivers of Blue Nile variability are:
- The identification of a tropical Indian Ocean influence on East Africa summer rains, in which low pressure in the Indian Ocean enhances the flow of moist near-surface westerly winds across the Congo Basin and into Eastern Ethiopia.
- Complementary work on the westerly wind influence that identifies potential predictors in surface or near-surface pressure anomalies in the tropical Atlantic and the Arabian Peninsula, or the gradient between the two.
- Evidence that the southern hemisphere subtropical high pressure centers—the Mascarene in the Indian Ocean and the St. Helena in the Atlantic Ocean—influence the inflow of water vapor to the region.
- Studies showing that the strength and location of the Tropical Easterly Jet (TEJ) and, to some extent, the African Easterly Jet (AEJ) modify convection in the region. Links between the TEJ and the Quasi Biennial Oscillation (QBO) have raised the possibility of QBO as a predictor of Blue Nile flow.
- Perhaps most prominently, extensive study of the influence that the El Niño Southern Oscillation (ENSO) has on summertime Ethiopian rainfall, both synchronously and with a time lag. The ENSO influence is preeminent in operational statistically-based prediction systems, and its impact is clear in almost any analysis of rainfall variability (e.g., Figure 5). During the summer rains, El Niño events are associated with below average rainfall; this is the opposite of the well-known influence ENSO has on the eastern Horn of Africa and Equatorial East Africa during the October-December short rains. The mechanism of the summertime ENSO influence, however, is still debated, and could be multifaceted. Influences of ENSO on subtropical highs, on water vapor advection, and on the TEJ and AEJ have all been proposed.
Figure 5: annual rainfall anomaly (mm) associated with El Niño relative to non-El Niño years. Precipitation data are from CHIRPS for the period 1981-2014. El Niño years are defined in terms of Niño 3.4 anomaly.
This partial list of drivers of variability points to the potential for teleconnection-based prediction in the region. But it also highlights the complexity of interacting climate dynamics that affect the Blue Nile basin (Figure 6). These influences also appear to vary over the course of the season. Berhane et al. (2013) find that the influence of ENSO and of Indian Ocean predictors tends to be strongest late in the rainy season and non-significant early in the season. Further, the influences are non-stationary in time. The ENSO teleconnection, for example, which is central to many prediction systems, was stronger prior to 1976 than afterwards, and recent warming in the western Pacific and Indian Oceans has had a substantial impact on rainfall across East Africa.
Figure 6: schematic map of some of the leading proposed influences on the Ethiopian summer rains. ENSO = El Niño Southern Oscillation, AEJ = African Easterly Jet, TAH = Tropical Atlantic High, SHH = St. Helena High, green arrow = Mediterranean Sea water vapor import, AL = Arabian Low, SAM = South Asian Monsoon, TEJ = Tropical Easterly Jet, IOSST = Indian Ocean SST, MH = Mascarene High.
Existing seasonal forecast models
The importance of rainfall in East Africa has inspired many efforts at seasonal prediction, including a number that are directly relevant to Blue Nile flow forecast. The majority of these efforts were undertaken for research and are not currently operational, while a smaller number are available as operational forecast systems. We divide these efforts into statistically-based forecast models, which are based on observed relationships between rainfall (or river flow) and time lagged large scale predictors in the climate system, and dynamically-based forecast systems, which employ global climate models to forecast atmospheric conditions and surface meteorology at lead times of weeks to months.
Most of the published statistical models of Ethiopian summertime rainfall–the season that matters most for Blue Nile flows–derive their skill all or in part from associations with ENSO and/or Indian Ocean surface conditions (Table 1). Gissila et al. (2004) leveraged these teleconnections to generate a nine predictor linear regression model in which summertime Ethiopian rainfall is positively associated with western Indian Ocean SST and negatively associated with SST in the tropical Pacific Ocean (Niño 3.4 region). This is consistent with empirical relationships and proposed mechanisms described above. Korecha & Barnston (2007) take a slightly different approach, using the evolution of Niño 3.4 SST and tropical Atlantic Ocean SSTs in the months leading up to the rainy season as predictors, alongside the late spring Niño 3.4 SST. Their predictive regression model indicates that springtime cooling in the Niño 3.4 region is associated with more precipitation, perhaps because this metric captures a trend away from El Niño (or a trend into La Niña) conditions in the months leading up to the onset of rains.
Nicholson (2014) included atmospheric fields in her identification of predictors. This led to a three variable linear regression model in which summertime rains in East Africa are a function of zonal wind strength in the TEJ exit region, SST gradients in the tropical and subtropical Pacific Ocean, and tropical Indian Ocean SLP. Segele et al. (2015) cast a wide net for predictors, generating an ensemble of multivariate regression models from a family of potential atmospheric and SST predictor fields. The Ethiopian National Meteorological Agency use a statistical analog approach based on ENSO indices. Historical analog years are selected based on similarity in a family of ENSO indicators diagnosed both from observed and forecast ENSO conditions for the coming rainy season. These analog years are then applied to forecast seasonal total rainfall terciles for homogenous regions across Ethiopia and also to forecast second order properties of the rainy season such as onset and cessation of rains and probability of extreme events (Korecha & Sorteberg, 2013).
A number of global dynamically-based seasonal forecast systems are also available for application to Blue Nile forecast. The raw output of some of these dynamically-based systems are publicly available (notably all models participating in the North American Multi-Model Ensemble (NMME); Kirtman et al. 2014), and all centers generate products that can be incorporated into value-added analyses by operational forecasting organizations. Analysts with the Famine Early Warning System (FEWS), for example, generate outlooks that take into account dynamically-based predictions from NMME, the European Centre for Medium-range Weather Forecasts (ECMWF), and the United Kingdom Meteorological Service (Met UK), along with ancillary climate data and expert opinion. The Greater Horn of Africa Climate Outlook Forum (GHACOF) generates probabilistic consensus forecasts by merging multiple dynamically-based and statistical forecasts, based heavily on input from participating national meteorological agencies and the IGAD Climate Prediction and Application Center (ICPAC). The International Research Institute for Climate and Society (IRI) produces a widely used forecast based on bias corrected and weighted ensembles of dynamically-based forecast systems.
Here we examine as many of these statistically and dynamically-based forecasts as we had information and time to pursue, both to see how the models have performed in recent years and to provide an outlook for the 2018 rainy season. For statistical models that are not operational we adopted the predictors and model structure from the literature but refit all coefficients based on a consistent suite of datasets, updated to as recently as the data allowed. For operational systems that have issued their own 2018 predictions we simply report the publicly issued forecasts.
Models used in this study
The statistical models used in our analysis are listed in Table 1. In all cases we used June-September total precipitation from CHIRPS as the predictand. Models were fit using a combination of predictors from NCEP/NCAR reanalysis product and NMME lag-0 forecast products (forecasts for May, made in May). The NMME lag-0 fields were used for any May predictors required for 2018 June-September prediction, since latency in the reanalysis product makes it impossible to use May reanalysis fields for an operational prediction that is issued by the beginning of the rainy season. Leave-one-out cross-validation was used when fitting these models to optimize for out-of-sample predictive accuracy. The period of retrospective analysis was 1982-2017, the longest record for which we could obtain output from both statistical models and dynamically-based prediction systems. Recognizing that teleconnections might have shifted over this 36 year period, we also fit one model using the shorter 2001-2017 record in order to see how changes post-2000 might influence model skill (Blum-2 in Table 1). Our analysis of dynamically-based forecast systems is limited to models in the NMME ensemble, as those were the only models for which we could easily obtain both historical and operational forecasts. The years 2011 and 2017 were not available for all NMME models included in the analysis, so the evaluation period was restricted to 1982-2010 plus 2012-2016. Precipitation from NMME models was bias corrected to CHIRPS prior to evaluation.
Table 1: Statistical models used in this study. P_JJAS=Sum of precipitation for June, July, August and September; SST=Sea Surface Temperature; All variables (predictors and predictand) were standardized (subtract mean and divide by standard deviation 1982-2017). Note that in applying previously published models we retain their structure and list of candidate predictors, but coefficients and, in some cases, selected variables are different from the originally published versions.
|Model/Paper||Model Structure||Predictor variables|
|Gissila et al. (2004)||Multivariate linear regression||March, April, May SST for tropical western Indian Ocean, 10ºS-10ºN, 50-70ºE, tropical eastern Indian Ocean, 10ºS-0,90º-110ºE and Nino3.4|
|Korecha & Barnston (2007)||Multivariate linear regression||The difference of May minus February–March SSTs over the south Atlantic, 30°S–40°S, 30°W–15°W|
|The difference of May minus the February–March Niño-3.4 SST, 5°N–5°S, 170°W–120°W|
|May Niño-3.4 SST, 5°N–5°S, 170°–120°W|
|Nicholson (2014)||Multivariate linear regression||May u 200 hPa, 30ºE-50ºE and 0º -10ºN|
|May difference in SST:
(170º -265ºE, 5ºS – 5ºN) – (137ºE-160ºE, 18ºN – 28ºN)
|May sea level pressure (SLP) for 80º -90ºE, 5º 15ºN|
|Segele et al. (2015)||Mean of 11
Multivariate linear regression models (Fig.5 in Segele et al. 2015)
|20 predictors from Table 1 in Segele et al .(2015)|
|Wu – this study||Linear Inverse Model||May, June, July, Aug and Sep EOFs of Local Precipitation, Tropical Pacific SST, Tropical Atlantic SST, Indian Ocean SST, North Pacific SST, North Atlantic SST,
Southern Ocean SST, Global 850 hPa and 200 hPa Geopotential height
|Alexander – this study||Principal Component Regression||First two principal components of May Tropical Pacific Ocean SST, March-Feb Central Indian Ocean SST, April Tropical Atlantic SST, and Indian Monsoon SLP Index|
|Blum-1 – this study||Multivariate linear regression||May SST for tropical western Indian Ocean, 10ºS-10ºN, 50ºE-70ºE
May sea level pressure (SLP) for 80ºE – 90ºE, and 5ºN – 15ºN
May difference in SST: (170ºE – 265ºE, 5ºS – 5ºN) – (137ºE- 160ºE, 18ºN – 28ºN)
Southern Europe air temperature at 100hPa, 10ºE-15ºE, 45ºN-50ºN
North Arabian Sea March meridional temperature flux at 925 hPa, 60ºE-65ºE, 20ºN-25ºN
|Blum-2 – this study||Multivariate linear regression (fit using 2001-2017 data)||Eastern Indian Ocean zonal geopotential height flux at 300 hPa, 80º-85ºE, 5ºS-0º
Arabian Sea meridional moisture flux at 925 hPa, 60º-65ºE,10º-15ºN
North tropical Pacific SST, 171ºE-175ºE, 16ºN-18ºN
Southeastern Europe zonal temperature flux at 100hPa, 25ºE-30ºE, 40ºN-45ºN
Historical model performance for prediction of June-September precipitation is summarized in Table 2, corresponding to the hindcasts plotted in Figure 7. This summary is intended to give a general perspective on performance rather than a detailed or exhaustive assessment of model behavior. We consider the anomaly correlation coefficient (ACC) as a measure of linear relationships between predictions and observations independent of mean and variance, root mean square error (RMSE) of the anomalies to quantify errors in variance, and bias. We also quantify the Hit Score for predictions, which is calculated by dividing the observational record into three terciles (below normal, near normal, and above normal) and calculating the fraction of predictions that fell into the correct category over the period of analysis. To focus on extremes, we include both a standard Hit Score and a Hit Score for Extremes, which considers only the hit rate for top and bottom tercile events. All scores are for cross-validation out-of-sample predictions, rather than for model fit.
Figure 7: 1982-2017 hindcasts for June-September Blue Nile basin precipitation. Grey lines are statistical models, with the solid black line showing the ensemble average of statistical models applied in this project. Blue lines are NMME forecasts initialized in May of each year–each light blue line is the ensemble average of a single model and the dark blue line is the full NMME ensemble average for all models that had complete data records. Dashed black line is CHIRPS June-September precipitation anomaly averaged across the Blue Nile basin.
Table 2: Evaluation of statistical and dynamically-based hindcasts for the period 1982-2017 (for NMME models only 1982-2010 and 2012-2016 were available for all models).
|Models||ACC||RMSE (mm)||BIAS (mm)||HitScore||HitScore_Extremes|
|Alexander – this study||0.70||52.2||-3.5||0.61||0.70|
|Blum-1 – this study||0.70||51.7||1.0||0.58||0.65|
|Blum-2 – this study||0.43||67.1||1.1||0.39||0.48|
|Wu – this study||0.66||55.7||-1.4||0.64||0.70|
|Mean Statistical Model||0.66||56.7||-2.8||0.58||0.65|
Table 3: Evaluation of statistical and dynamically-based hindcasts for the period 2001-2017 (for NMME models only 2001-2010 and 2012-2016 were available for all models).
|Models||ACC||RMSE (mm)||BIAS (mm)||HitScore||HitScore_Extremes|
|Alexander – this study||0.77||42.3||-15.2||0.67||0.73|
|Blum-1 – this study||0.49||58.7||1.3||0.39||0.36|
|Blum-2 – this study||0.45||60.5||-9.1||0.33||0.36|
|Wu – this study||0.69||48.5||6.1||0.61||0.64|
|Mean Statistical Model||0.72||51.3||13.0||0.44||0.55|
Our evaluation of NMME members confirms previous work with ensembles of dynamically-based prediction systems, in that performance is quite mixed across different models. The NMME ensemble average prediction, however, shows predictive skill that is comparable to many of the statistical models, and the NMME hit rate is as good as the top performing statistical models. Table 2 shows large bias values for some NMME models prior to bias correction, but this does not affect the application of the models to study anomalies. We note that these are direct NMME forecasts of precipitation in the GERD watershed. We did not consider “hybrid” approaches here, in which dynamically-based forecast systems are used to predict the predictors for a statistical model (e.g., Shukla et al. 2014, Gleixner et al. 2017), and which represent another powerful way to make use of dynamically-based models.
For the statistical models, we find that skillful predictions can be obtained using a number of different model structures and sets of predictors. All models rely to some extent on associations with surface conditions in the Tropical Pacific and/or Indian Ocean, reflecting the known importance of lagged ENSO connections and Indian Ocean conditions to the region. Skillful models also generally include some combination of predictors that capture the influence of Tropical Atlantic conditions or of atmospheric predictors that capture synoptic dynamics affecting East Africa, though the combination of these predictors varies by model. This is consistent with the recent work of Nicholson (2014) and Segele et al. (2015) who showed that diversifying predictors beyond the traditional ENSO and Indian Monsoon zones can enhance predictive skill. The fact that skill can be obtained using different combinations of these predictors makes it difficult to apply these prediction-oriented models to advance fundamental understanding of the drivers of summertime rainfall variability. These mechanisms, including the relationship between correlated predictors in these models, is the subject of ongoing research in the climate community. The models included here show some range in statistical performance. In general, the models that were customized for Blue Nile basin rainfall and developed using more recent data records show slightly higher skill for this application. A prediction based on the mean forecast of all statistical models fares well in both evaluation time periods.
While the 1982-2017 baseline has the advantage of offering a 36 year record for model evaluation, we do not necessarily expect that model skill is stationary over this full period. Indeed, the strength of various teleconnections to East Africa is known to vary as a function of climate variability, and over the past 37 years it is also possible that global warming has altered climate dynamics relevant to prediction. For this reason we also evaluate all models over the shorter 2001-2017 period (Table 3). Evaluation for this shorter record is not as statistically robust, but it offers a view of how well models are doing under more current conditions. All of the previously published statistical models show poorer performance for 2001-2017 than they do for the whole record. This is not surprising, as these models were developed without using the most recent years that are now available; we recalculated the coefficients using these recent years, but we did not revisit model structure. The magnitude of the drop-off is rather large, however, which is consistent with the observation that correlation with previously reliable predictors in the Pacific Ocean has been inconsistent since 2000. Several of the models we developed specifically for this study maintain their level of skill when tested against this more recent period. NMME models show some changes in skill for the 2001-2017 period, but these changes show no evidence of being systematic, at least within the limits of this cursory analysis. In sum: statistical approaches provide meaningful skill for seasonal forecast of Blue Nile basin rainfall, though there is considerable unexplained variance for all modeling approaches. The NMME models vary widely in their performance, but the NMME ensemble average offers respectable performance relative to many statistical approaches.
Blue Nile flow
Both the statistically and dynamically-based models described above predict rainfall. To convert a rainfall forecast to a Blue Nile streamflow forecast we adopted a variation of the Water Balance model (WatBal; Yates 1996). It takes monthly precipitation (P), and climatological monthly mean temperature (T) and diurnal temperature range (DTR) to produce monthly streamflow within a river basin. CRU TS data are used for climatological T and DTR. CenTrends and CHIRPS data are used for precipitation. Note that in order to take the predicted JJAS total precipitation as the updated inputs to produce the predicted streamflow, the model is calibrated using the disaggregated historical JJAS monthly P and the climatology of T and DTR. Climatological rainfall was used for Oct-Dec of 2018, as there is little rain in these months.
The calibrated WatBal model offers strong performance in the historical record. The monthly time series show a correlation of 0.96 (p-value <.0001) between the observation and calibration (Figure 6A). For interannual variability on total flow the correlation is 0.77 (p-value <.0001). The scatter plot of monthly observation and calibration also aligns with the 45 degree line (Figure 6B). However, in general the calibrated streamflow shows an overestimation during early months of the calendar year while an underestimation during the later months (Figure 6C). Accumulatively, the calibrated streamflow is approximately 4% higher than the observed streamflow annually. Flow estimates generated by applying precipitation hindcasts to the hydrological model are shown in Figure 6D.
Figure 8: (A) comparison of observed streamflow at El Diem and simulated calibrated streamflow from WatBal, in units of million cubic meters (Mm3); (B) scatter plot of observed monthly streamflow at El Diem (x-axis) and calibrated streamflow from WatBal (y-axis); (C) monthly average streamflow based on observations at El Diem and calibrated streamflow from WatBal. Data for all plots is 1965-2009; (D) hindcasts of June-Dec total Blue Nile flow at the GERD site (Mm3) generated by applying statistical (grey lines) and dynamically-based (blue lines) precipitation hindcasts to the WatBal model. Dashed black line in (D) is the result of historical WatBal simulation driven with CHIRPS rainfall.
The outlook for 2018
The outlook for 2018 is highly informed by the fact that we are currently in the waning months of a La Niña event, which is expected to fade to neutral conditions that will persist through the summer rains (Figure 8). Since most droughts in the Blue Nile basin occur during El Niño events, current conditions and the majority of projections through the season suggest that drought is unlikely to occur. There is, however, the potential for El Niño conditions to emerge before the end of the season, which might affect late season rains in a manner that the statistical models considered here do not directly capture. Many factors play into climate variability in the Blue Nile basin, but the absence of El Niño provides a strong precondition towards average or above average Blue Nile flows.
Figure 9: Oceanic Niño Index forecast for 2018, generated by NOAA CPC and the International Research Institute (IRI) and issued on May 14, 2018.
As shown in Figure 2, both statistical and dynamical models suggest that there is a strong likelihood of average to above average rainfall in the Blue Nile basin in the 2018 summer rainy season. The predictions are not uniform, however. There is a distribution in statistical models that includes three models with a median prediction that is in the upper tercile of the historical record, four models that are in the middle tercile, and one that is in the lower tercile (Figure 10). For NMME, predictions for seven of the eight models we were able to obtain have a median forecast in the top tercile. There is, however, considerable spread in the absolute forecast of the NMME models.
Figure 10: Boxplots illustrating range of NMME ensemble predictions and statistical models with re-sampled error added back in. Dashed lines show below/above normal conditions. Statistical boxes represent range of 2018 forecast plus 100 resampled errors (from 1982-2017 leave-one-out predictions).
The spread shown in Figure 10 is a combined result of differences in rainfall variability between models–GEOS_S2S, for example has a large interannual standard deviation in precipitation, and Figure 10 shows bias corrected but not standardized forecasts–and of differences in the forecast relative to each model’s historical record. Standardized anomalies range from 0.5 to > 2.0 standard deviations above the historic mean. We chose not to standardize when presenting results because the majority of models have variability that is smaller than CHIRPS observed interannual variability. This means that for a wet forecast like 2018, presenting the absolute rather than variance-adjusted anomalies represents a conservative error when assessing the potential volume of excess flow predicted to be available. Ensemble spread across realizations of each NMME model is also substantial, with select ensemble members of several models yielding below average precipitation forecasts. The uncertainty in both statistical and dynamically-based models is clear in the smoothed histograms shown in Figure 11, which indicate that both families of forecasts include some members that allow for below average conditions to develop in 2018, even if it is a low probability. Figure 12 shows the same smoothed distribution for streamflow estimates.
Figure 11: Smoothed histograms showing climatology (1982-2017) compared to statistical and NMME predictions for 2018 June-September rainfall. Dashed lines show below/above normal conditions. Note: climatology (n=37),NNME (n=8),statistical (n=8)
Figure 12: As in figure 11, but for June-December streamflow at the site of the GERD. Note: climatology (n=37),NMME (n=8),statistical (n=8)
These precipitation forecast results can be compared to those issued by operational forecasting institutions. The dynamically-derived IRI forecast of May 15, 2018, for example, shows somewhat elevated probability of high rainfall over much of the Blue Nile basin (Figure 13). The NOAA Climate Prediction Center (CPC) categorical forecast, calculated from the full, unweighted NMME ensemble, also shows a slightly elevated probability of above average rainfall in Ethiopia (Figure 14). ECMWF June-August precipitation forecasts are for normal rainfall conditions in the Blue Nile region. The GHACOF and Ethiopia NMA forecasts were not available at time of writing, but we will update to include them when they go online.
Figure 13: IRI July-September precipitation forecast, derived from weighted analysis of dynamically-based forecast systems. Red box indicates approximate location of the GERD watershed.
Figure 14: CPC categorical predictions for July-Sep precipitation in the May NMME ensemble.
While seasonal precipitation and hydrological forecasts of the Blue Nile River carry substantial uncertainties, the relative consistency of the 2018 prediction across models of different origin and structure provides some confidence that there is a high probability of average to above average flow in the coming season. The sampling of models considered here suffer from inconsistent performance in recent years, sensitivity to calibration period (for statistical models) and to uncharacterized spatial biases (for dynamically-based systems), and the inherent limitations that come when forecasting an imperfectly understood system. Nevertheless, the consensus outlook is consistent with (and, indeed, driven in large part by) the understanding that drought in this region usually coincides with El Niño events, and we are currently experiencing a fading La Niña, with ENSO neutral conditions likely in the coming months. We do note that the ENSO teleconnection to the Blue Nile basin is not fully understood and appears to be inconsistent over time. This contributes to forecast uncertainty, and is an important area of ongoing research.
This exercise has been motivated in large part by the fact that Ethiopia will soon begin to fill the reservoir of the Grand Ethiopian Renaissance Dam. This is a major hydrological activity with considerable economic, social, and political implications. Decisions regarding the timing and rate of filling clearly depend on many factors that fall beyond the scope of seasonal forecast. Nevertheless, we feel that it it is valuable for Parties involved in filling decisions to have a common and realistic set of expectations for water availability in each year of the filling period. Consensus seasonal hydrological forecasts can inform those expectations on a year by year basis, perhaps easing one source of tension as riparian countries engage in the shared challenge of managing a transboundary waterway under rapid development.
Berhane, F., Zaitchik, B., & Dezfuli, A. (2014). Subseasonal analysis of precipitation variability in the Blue Nile River Basin. Journal of climate, 27(1), 325-344.
Gissila, T., Black, E., Grimes, D. I. F., & Slingo, J. M. (2004). Seasonal forecasting of the Ethiopian summer rains. International Journal of Climatology, 24(11), 1345-1358.
Gleixner, S., Keenlyside, N. S., Demissie, T. D., Counillon, F., Wang, Y., & Viste, E. (2017). Seasonal predictability of Kiremt rainfall in coupled general circulation models. Environmental Research Letters, 12(11), 114016.
Kirtman, B. P., Min, D., Infanti, J. M., Kinter III, J. L., Paolino, D. A., Zhang, Q., … & Peng, P. (2014). The North American multimodel ensemble: phase-1 seasonal-to-interannual prediction; phase-2 toward developing intraseasonal prediction. Bulletin of the American Meteorological Society, 95(4), 585-601.
Korecha, D., & Barnston, A. G. (2007). predictability of June–September rainfall in Ethiopia. Monthly weather review, 135(2), 628-650.
Korecha, D., & Sorteberg, A. (2013). Validation of operational seasonal rainfall forecast in Ethiopia. Water Resources Research, 49(11), 7681-7697.
Nicholson, S. E. (2014). The predictability of rainfall over the Greater Horn of Africa. Part I: Prediction of seasonal rainfall. Journal of Hydrometeorology
Nicholson, S. E. (2017). Climate and climatic variability of rainfall over eastern Africa. Reviews of Geophysics.
Segele, Z. T., Richman, M. B., Leslie, L. M., & Lamb, P. J. (2015). Seasonal-to-interannual variability of ethiopia/horn of Africa monsoon. Part II: Statistical multimodel ensemble rainfall predictions. Journal of Climate, 28(9), 3511-3536.
Shukla, S., Funk, C., & Hoell, A. (2014). Using constructed analogs to improve the skill of National Multi-Model Ensemble March–April–May precipitation forecasts in equatorial East Africa. Environmental Research Letters, 9(9), 094009.