Managing climate risks in agriculture and hydropower by coupling prediction

and multi-purpose reservoir models in the Blue Nile Basin, Ethiopia

By

Sarah O. Alexander

A thesis submitted in partial fulfillment of

the requirements for the degree of

Master of Science

(Civil and Environmental Engineering)

at the

UNIVERSITY OF WISCONSIN-MADISON

2018

ii

ACKNOWLEDGEMENTS

This work would not be possible without the guidance, collaboration, and support of Dr. Paul

Block. His commitment to research and advising has led to both personal and professional growth

through this research. Committee members Dr. Daniel Wright and Dr. Anita Thompson as well as

anonymous reviewers of a manuscript in preparation provided insight and critical feedback that

has elevated the caliber of the work.

The material presented in this thesis is based upon work supported by the National Science

Foundation under Grant No. 1545874.

Finally, I would like to acknowledge many organizations and individuals for their collaboration,

insight, and feedback, including: National Meteorological Agency of Ethiopia, Ethiopia Ministry

of Water Resources, Abay Basin Authority, Girmachew Addisu, Dr. Shu Wu, PIRE colleagues at

the University of Connecticut, Water Systems and Society research colleagues, the Water

Resources Engineering group, friends, and family. Their support and engagement have greatly

contributed to the success of the work.

iii

ABSTRACT The Blue Nile Basin in Ethiopia is a critical water resource for the nation and continent, subject to

spatial and temporal variations in climate that adversely impact local communities. Extreme

conditions, e.g. drought and flood, threaten livelihoods causing direct hydrologic, economic, and

social implications. To reduce vulnerability and enhance economic security in the region,

predictive information is coupled with operational reservoir models to issue advance prediction of

dominant rainy season conditions and inform sectoral decision-making.

Advance predictions of seasonal precipitation may provide information to aid water resource

management decisions in various sectors. Yet, a disconnect between the spatial scale upon which

skillful predictions are issued and the sectoral decision-making scale renders current predictive

information inadequate in many cases. This work explores season-ahead precipitation prediction

skill for a local region in the Blue Nile basin, Ethiopia, to better understand how model structure

may serve to provide more skillful and valuable predictive information to the end-user. Statistical

downscaling of global dynamic and regional empirical models as well as development of a high

resolution, locally-tailored statistical model indicate that model structure and prediction skill are

inextricably linked. Statistical approaches specific to the local region show higher prediction skill

at the sectoral decision-making scale compared with dynamic approaches, offering the potential to

aid local communities in many regions that are currently vulnerable to highly variable spatial and

temporal precipitation patterns.

Predictive hydroclimate information may be further enhanced by integrating with sectoral

models, to provide actionable information for communities by explicitly addressing climate

risks. In this work, local-scale, statistical streamflow forecasts are coupled with a multi-purpose

reservoir model to explore benefits and tradeoffs to hydropower and agriculture production for

the Finchaa – Amerti system in the Blue Nile basin of Ethiopia. Strategies to meet agriculture

demand and maximize hydropower result in increased hydropower production and agriculture

water allocation using this framework. The resulting increase in resilience and decrease in

vulnerability across sectors provides a concrete example of how hydroclimate prediction can be

translated to actionable information to guide water resources management and bridge the existing

disconnect between prediction and decision-making scales.

iv

TABLE OF CONTENTS

Chapter 1: Introduction and Context 1

Background: Blue Nile basin, Ethiopia 1

Case Study Regions 2

Local Scale Sectoral Models 5

Chapter 2: Local-scale Precipitation and Streamflow Prediction 9

Data and Forecast Verification Metrics 9

Data 9

Forecast Performance Metrics 11

Downscaling Global Dynamic and Regional Empirical Models 12

Dynamic Models 12

Empirical Cluster Model 14

Local Statistical Prediction Model 15

Model Framework 15

Model Evaluation and Comparison 19

Chapter 3: Coupled Forecast and Reservoir Models 25

Application of Precipitation Predictions for Reservoir Management 25

Framework for Coupling Seasonal Precipitation and Reservoir Volume 25

Using the Local Statistical Model to Predict Reservoir Storage 27

Coupling Streamflow Predictions with Operational Reservoir Models 28

Model of the Finchaa – Amerti System 28

Prediction-coupled Operational Reservoir Model 29

Measures for Assessment of Model Performance 31

Evaluation and Prediction-coupled Reservoir Model Performance 32

Opportunity for Future Re-operation of the Finchaa – Amerti System 38

Chapter 4: Discussion and Conclusions 40

References 45

1

CHAPTER 1: INTRODUCTION AND CONTEXT Background: Blue Nile basin, Ethiopia

The Blue Nile basin (BNB) in Ethiopia is a critical water resource for the region and continent,

serving as the largest tributary to the Nile River by providing over 70% of downstream flows at

the Aswan Dam (Conway 2000, 2005; Siam and Eltahir 2017). A majority of the basin experiences

strongly seasonal precipitation, with most falling during the Kiremt rainy season, June – September

(JJAS; Yates and Strzepek 1998). The relative wealth of water resources in the basin affords the

potential for economic gains through irrigated agriculture and hydropower. Current agricultural

practices, however, are primarily rain-fed and represent the dominant sector in Ethiopia; thus,

precipitation is integral to sustaining the economy and livelihoods. Yet precipitation across the

region remains subject to high spatial and temporal interannual variability (Conway 2000, 2005;

Segele and Lamb 2005; Zhang et al. 2016). Previous work indicates that these interannual

variations result from movement of the inter-tropical convergence zone (ITCZ), regional climate

interactions, topographic influences, and large-scale climate teleconnections, such as the El-Nino

Southern Oscillation (ENSO) (Bekele 1997; Block and Rajagopalan 2007; Camberlin 1997;

Conway 2000; Diro et al. 2009; Gissila et al. 2004; Korecha and Barnston 2007; Segele et al. 2009;

Segele and Lamb 2005; Zhang et al. 2016). Teleconnections to the Indian Ocean and other high

pressure systems are also cited as contributing factors to interannual variation (Black 2005; Folland

et al. 1986; Segele and Lamb 2005). The strong spatial and temporal variability in hydroclimate

often results in relatively proximal local areas experiencing varying amounts or timing of

precipitation throughout the season. Thus, hydroclimate conditions averaged at the regional scale

– for which predictions are most often available – may not always be representative of the local,

decision-making scale.

Season-ahead predictions from dynamic and statistical models have been previously investigated

for regions in East Africa. Dynamic model approaches have largely focused on precipitation

prediction at the regional scale, using either regional climate models (RCMs) or statistical methods

to downscale global climate models (GCMs) for East Africa, however some studies indicate little

additional value with RCM downscaling (Cheneka et al. 2016; Diro et al. 2012; Nikulin et al. 2017;

Ogutu et al. 2017; Shukla et al. 2014a). For precipitation across Ethiopia specifically, dynamic

2

model predictions generally exhibit modest to poor skill (Gleixner et al. 2017; Jan van Oldenborgh

et al. 2005; Jury 2014; Shukla et al. 2016). Statistical model approaches for precipitation prediction

across East African locations typically employ oceanic and atmospheric climate variables and

teleconnections directly (Camberlin and Philippon 2002; Funk et al. 2014; Hastenrath et al. 2004;

Mwale and Gan 2005; Ntale et al. 2003; Philippon et al. 2002; Yeshanew and Jury 2007). Some

statistical approaches use analog approaches based on large-scale climate variables for

precipitation prediction (e.g. Obled et al. 2002). Statistical modeling approaches applied

specifically in Ethiopia include national-level (e.g. Korecha and Barnston 2007) and sub-national

scale approaches to capture heterogeneity of moisture transport processes and precipitation,

particularly in the Ethiopian highlands (Bisetegne et al. 1986; Diro et al. 2011a; Eklundh and

Pilesjö 1990; Gissila et al. 2004; Zhang et al. 2016). The National Meteorological Agency (NMA)

of Ethiopia has issued sub-national precipitation forecasts across all of Ethiopia, conditioned

primarily on the state of ENSO, since 1987 (Korecha and Sorteberg 2013). Diro et al. (2011b) use

Pacific, Indian, and Atlantic sea-surface temperature anomalies as predictors with multiple linear

regression and linear discriminant analysis techniques. Still others focus at the BNB scale, using

nonparametric, statistical methods to generate ensemble precipitation forecasts (Block and

Rajagopalan 2007), or k-means statistical methods to optimally determine clustered regions and

provide tailored precipitation predictions for each cluster (Zhang et al. 2016). Even though these

empirical models demonstrate skill at the cluster level, regressing predictions to higher resolution

(approximately 11 km) has been shown to cause a significant deterioration in skill (Zhang et al.

2017). Thus, even though numerous sub-national to East African regional scale seasonal

precipitation predictions have been developed, all with modest skill, a connection to expected

conditions at the local scale is practically nonexistent, yet could serve to support agricultural,

energy, environmental, or other sectoral decisions.

Case Study Regions

Two local regions within the BNB, Ethiopia, are selected for local hydroclimate prediction model

assessment based on vulnerability to climate variability and potential to improve local economic

security (Figure 1). The regions exhibit a strongly seasonal distribution of precipitation, with

precipitation primarily across the JJAS months, during which cereal and subsistence crops are

planted (e.g. long and short cycle grains, coffee, pulses, oilseeds). Portions of each region also

3

benefit from an installed reservoir, allowing for irrigated agriculture during a second cropping

season in October – May, and potentially increasing the annual agricultural yield and household

income through cash crops (e.g. short-cycle grains, corn, sugar cane, coffee; Eriksson 2013). Thus,

predictive information has the potential to enhance rain-fed and irrigated agricultural management

decisions.

Figure 1. Map of Blue Nile (Abay) basin, Ethiopia with the Koga and Finchaa basins shown

in red and Medium Blue Nile Region (MBNR) outlined.

The Koga basin is located in the Mecha Woreda, West Gojam Zone of the Amhara state,

approximately 35 km southwest of Bahir Dar. The basin transitions from mountainous areas

upstream to relatively flat areas downstream; elevations across the basin range from 3000 to 1800

meters above sea level (m.a.s.l.) (Gebrehiwot et al. 2010; Reynolds 2013). Annual precipitation

averages 1419 mm, with approximately 75% coming in the JJAS season, however JJAS

precipitation varies notably year to year (Figure 2a; coefficient of variation of 12%). The Koga

Dam and Irrigation Site was constructed in 2011 to provide large-scale irrigation to farming

communities downstream in an effort to increase economic security in the region through irrigated

agriculture during the dry season. Located on the Koga River, a tributary to the Gilgel Abay River

that flows to Lake Tana, the site includes a dam and 83.1 million cubic meter (Mm3) capacity

reservoir to capture river inflow, precipitation, and runoff from a 220 km2 catchment area. A

system of canals carries water from the reservoir to provide irrigation to approximately 7000 ha of

Medium Blue Nile Region

4

agricultural land (Eriksson 2013; Reynolds 2013). In the Koga basin, over 70% of the river flow

is a result of precipitation during the JJAS rainy season, thus the ability to capture water for a

second crop in the dry season provides significant benefits to the surrounding communities

(Reynolds 2013).

Figure 2. Timeseries of total annual precipitation (black) and JJAS season precipitation

(blue) for the Koga (a) and Finchaa (b) regions. Long term mean JJAS precipitation amount

is shown in red; coefficient of variation is 12% (a), 9% (b).

The Finchaa watershed is located in the southern portion of the Abay basin and covers

approximately 1318 km2. The watershed is predominately rolling plateaus with elevations ranging

a

b

5

from 3100 to 2200 m.a.s.l. Average annual precipitation is 1442 mm, with 72% falling during the

JJAS rainy season; like Koga, Finchaa experiences notable inter-annual variability in rainy season

precipitation (Figure 2b; coefficient of variation of 9%). The Finchaa Dam was constructed in

1973 to foster economic growth in the region through hydropower and irrigated agriculture (Tefera

and Sterk 2008). Located on the Finchaa River, the dam is fed by the Finchaa and Amerti Rivers,

with a 1050.5 Mm3 reservoir. The hydropower plant has a 134 Megawatt (MW) installed capacity

from 4 turbines (Wilson 2007). Downstream releases are utilized at a sugarcane plantation and

factory, fisheries, and as a local water source, providing irrigation for over 20,000 ha of agricultural

land (Belissa 2016). Hydropower and agriculture demands are higher in the non-peak flow months

of October – May, with less water demanded during the JJAS rainy season, allowing for reservoir

filling (Belissa 2016). A dam on the Amerti River fills a 64.4 Mm3 reservoir, with diversions

released to the Finchaa reservoir providing additional water storage. In 2012, a dam was installed

on the Neshe River with a hydropower plant consisting of 97 MW total capacity (Wilson 2007).

While Neshe is now considered part of the Finchaa-Amerti system, releases from the Neshe

reservoir flow to the Finchaa River downstream of the reservoir. Thus, Neshe will not be

considered as part of this study.

Local Scale Sectoral Models

Season-ahead predictions of hydroclimate variables, such as precipitation or streamflow, have the

potential to aid water resource management in many areas, yet challenges remain, including skill

at the decision-making scale, prediction lead time, and cascading uncertainty. Multiple forecast

frameworks for season-ahead prediction, both statistical and dynamical, have been proposed in

recent studies (Block and Goddard 2012; Coelho et al. 2004; Goddard et al. 2001; Goddard and

Hoerling 2006; Kirtman et al. 2014; Korecha and Barnston 2007; Saha et al. 2006; Schepen et al.

2012; Shukla et al. 2014a; b). A common critique of these frameworks is their focus at a larger,

regional scale compared with the local scale at which water resource management decisions are

based (Gilles and Valdivia 2009; Pennesi 2007; Plotz et al. 2017). Research on communication

and uptake of forecasts repeatedly indicates that the disconnect between the scale of regional

predictions and local management decisions is a driving reason for lack of uptake, and further,

predictions at the local, decision-making scale have been shown to enhance value to farmers

(Gilles and Valdivia 2009; Plotz et al. 2017; Ziervogel et al. 2010). This motivates evaluation in

6

Chapter 2 of local-scale season-ahead prediction skill conditioned on regional dynamical and

empirical models as well as a local empirical model to understand scale effects, and essentially

address the question: are skillful hydroclimate forecasts at the local, sectoral decision-making

scale possible for enhanced value to the end-user?

For a hydroclimate forecast to have utility at the decision-making scale, model structures must be

capable of capturing small-scale variations in climate specific to the local region in addition to

larger-scale teleconnections (Plotz, Chambers and Finn 2017; Gilles and Valdivia 2009; Ziervogel

and Opere 2010). Dynamic models are state of the art, capturing complex climate dynamics and

physical interactions at the global scale, valuable for enhancing climate information in data limited

areas and for development of larger scale predictions. The accessibility of these forecasts as well

as the number of inter-connected climate variables they produce, lends them useful for many

applications (Kirtman et al. 2014; Shukla et al. 2016). Yet, biases, uncertainties, the inability to

capture smaller scale interactions, errors in parameter estimates or initial conditions, and other

discrepancies may result in unrealistic representations of local climate and poor prediction skill at

local scales (Block and Rajagopalan 2007; Lupo and Kininmonth 2013). Alternatively, statistical

prediction methods utilize climate teleconnections as predictors, and are often invoked for

prediction at smaller scales. While statistical models typically do not predict the full suite of

variables or well represent large spatial domains, as do dynamical models, the advantage of

explicitly using historical observations to condition relationships, without fully capturing complex

climate dynamics, may be in capturing smaller-scale inter-annual factors that modulate variability

on a local scale (Korecha and Sorteberg 2013; Zhang 2016). Therefore, a tradeoff may exist

between dynamic model structures, perhaps more suited to large-scale operational decision-

making, and statistical model structures, which may capture smaller-scale variations, depending

on the prediction scale desired. To evaluate, Chapter 2 compares local scale precipitation

predictions from downscaled dynamic and regional models and a local statistical model within the

Blue Nile basin.

Predictive information at the seasonal scale, linked with sectoral models offers the possibility of

altering planning and management decisions to avert climate variability risks for end-users.

Coupling of seasonal scale hydroclimate predictions with reservoir operations and management is

7

one example of how predictive information can be translated to actionable information with value

to users. Water resource management teams are increasingly interested in using predictive

information as a tool for mitigating risks from climate variability. For example, the U.S. Army

Corps of Engineers and other organizations are part of a research assessment titled Forecast

Informed Reservoir Operations (FIRO) to investigate whether climate forecast information can be

used to manage and guide reservoir operations in California (Jasperse et al. 2017).

Previous studies have coupled seasonal forecasts with reservoir models to assess the ability of

using predictive information to guide operational decisions, yet many of these studies rely on

dynamic models to provide the forecast information (Block et al. 2009; Crochemore et al. 2016;

Faber and Stedinger 2001). Dynamic models capture complex climate dynamics at the global scale

and often are easily accessible, rendering them useful for application to reservoir operations. For

example, Gong et al. (2010) develop a framework to use a set of a priori simulations that rely on

the historical record to represent future conditions, maintaining the rule curve structure while

adapting for smaller perturbations in the Delaware River Basin. Faber and Stedinger (2001) use

ensemble streamflow predictions from the National Weather Service and sampling stochastic

dynamic programming to optimize reservoir operations, while Crochemore et al. (2016) use Global

Climate Model (GCM) forecasts conditioned historical data to make best use of statistical and

dynamical methods.

Application of seasonal statistic predictions to reservoir management is limited, yet some

researchers have found that statistical predictions may increase reservoir performance (Gelati et

al. 2014; Georgakakos 1989; Golembesky et al. 2009; Hamlet et al. 2002). Maurer and Lettenmaier

(2004) explore the use of long-lead predictions for reservoirs on the main stem of the Missouri

River finding that economic benefits to hydropower are possible. Kim and Palmer (1997) use linear

regression methods to develop a seasonal snowfall runoff forecast for input into a Bayesian

Stochastic Dynamic Programming, demonstrating improved release patterns with the forecast.

Sankarasubramanian et al. (2009) generate a framework for using probabilistic seasonal forecasts

with water contracts for integrated management. More recently, LaMontagne and Stedinger (2018)

illustrate how statistically derived synthetic streamflow forecasts can be input to a simple reservoir

optimization to assess hydropower benefits. Given the recent success in using statistical methods

8

for prediction to inform reservoir operations and management, Chapter three seeks to understand

how an existing local-scale precipitation and streamflow forecast can be translated through a

reservoir model to answer the question: can local-scale statistical seasonal forecasts be coupled

with reservoir simulations to optimize operational strategies and provide actionable information

to inform water resources management?

Local scale precipitation predictions are used to issue an advance prediction of reservoir level at

Koga (due to data limitations in providing streamflow prediction directly), in order to provide

indication of how much water will be available in the dry season. Advance information of whether

the reservoir is expected to fill or not may prove helpful in decisions made prior to the dry season

including: amount of land to prepare, seed allocation, etc., with the opportunity for model output

to provide some indication of the degree to which the reservoir may not fill. Coupling of Finchaa

river streamflow predictions with a prediction-coupled reservoir model is used to understand the

benefits to competing water users under different operational strategies and explore options for

future re-operation.

9

CHAPTER 2: LOCAL SCALE PRECIPITATION AND STREAMFLOW PREDICTION

Data and Forecast Verification Metrics

Data

The Climate Hazards Group Infrared Precipitation with Stations (CHIRPS) product, at 0.05 degree

resolution, blends satellite information and gauge stations to provide a globally gridded

precipitation dataset for 1981 to the present (Funk et al. 2015). CHIRPS has been utilized in many

studies, and particularly in cases where observational records are poor (Funk et al. 2015; Shukla

et al. 2016). Validation and comparison of CHIRPS against similar products, including Climate

Forecast System (CFS) version 2, Climate Prediction Center Unified Gauge-based Analysis

(CPCU), European Centre for Medium-Range Weather Forecasts (ECMWF), and TRMM Multi-

satellite Precipitation Analysis (TMPA 3B42 RT7), indicates superior performance based on mean

absolute error, mean bias, and correlation metrics (Funk et al. 2015). In the BNB, CHIRPS data is

highly correlated with local precipitation gauge data (Figure 3) and is adopted as a representative

observational precipitation record for the available 37 years (1981-2017; spatially averaged across

the Koga (11.05 – 11.35 N, 37.05 – 37.35 E) and Finchaa (9.11 – 9.42 N, 37.00 – 37.30 E) regions).

10

Figure 3. CHIRPS and gauge station annual precipitation data for two stations near the Koga

case study region (correlation: a = 0.77, b = 0.43).

The North American Multimodal Ensemble (NMME) is a set of hindcast and real-time ensemble

forecasts (up to 9 months in advance) from a suite of Coupled Global Climate Models (CGCMs)

(Kirtman et al. 2014). NMME precipitation predictions across East Africa demonstrate more skill

than any single model and climatology (Shukla et al. 2016). Here, eight CGCMs providing 99 total

ensemble members are downscaled to examine seasonal precipitation prediction for the Koga

region in Ethiopia. Models used in analysis (and number of ensemble members) are as follows:

CMC1-CanCM3 (10), CMC2-CanCM4 (10), CCSM4 (10), GFDL-CM2pl-aer04 (10), GFDL-

CM2p5-FLOR-A06 (12), GFDL-CM2p5-FLOR-B01 (12), NASA-GMAO (11) and NCEP-CFSv2

(24). The NMME compares closely with other dynamic models, including ECMWF and GPCC

(Funk et al. 2015). NMME model hindcasts are available from the International Research Institute

for Climate and Society’s Data Library for 1982-2017 at a monthly, 1.0 degree resolution (Kirtman

et al. 2014). NMME seasonal total JJAS precipitation predictions for each year are averaged

spatially for each local region.

a

b

11

Statistical prediction models use global and local climate variables as predictors, including: sea

surface temperature, sea level pressure, surface air temperature, and precipitable water content.

All climate data were retrieved from the National Center for Environmental Prediction’s National

Center for Atmospheric Research (NCEP-NCAR) reanalysis project data in 2.5 by 2.5 degree

resolution for years 1948 to present (Kalnay et al. 1996).

Forecast Performance Metrics

Model hindcasts are evaluated with Pearson correlation, Hit Score, Extreme Miss Score, and Rank

Probability Skill Score metrics to compare performance (Regonda et al. 2006). Here, categories

are defined as below-normal (B), normal (N), and above-normal (A), with 33% of observations

falling into each category (climatological distribution).

The Hit Score (Barnston 1992) is calculated as follows (equation 1):

𝐻𝑖𝑡𝑆𝑐𝑜𝑟𝑒 = ∑,-./,,-.1,,-.23

𝑥100, (1)

where ∑𝐻𝑖𝑡7, 𝐻𝑖𝑡8,𝐻𝑖𝑡9 represents the sum of years where the median predicted ensemble value

and observed value from the hindcast fall into the same category, and 𝑛 represents the total number

of years in the timeseries. Misses represent years where the predicted category is incorrect, and are

particularly concerning when the prediction is off by more than one category (i.e. prediction of

above-normal when below-normal was observed, 𝑀𝑖𝑠𝑠7(9), or vice-versa), constituting an

“extreme miss” (equation 2):

𝐸𝑥𝑡𝑟𝑒𝑚𝑒𝑀𝑖𝑠𝑠𝑆𝑐𝑜𝑟𝑒 = ∑A-BB/(2),A-BB2(/)

3𝑥100, (2)

where ∑𝑀𝑖𝑠𝑠7(9),𝑀𝑖𝑠𝑠9(7) represents the number of years deterministic prediction values are

off by more than one category and 𝑛 represents the total number of years in the timeseries.

12

Rank Probability Skill Score (RPSS) provides the categorical skill of an ensemble forecast with

respect to a reference forecast (e.g. climatological distribution with A, N, B categories). The Rank

Probability Score (RPS) is the average of the squared difference between the cumulative

probability of the forecast and observations (equation 3):

𝑅𝑃𝑆 = ∑ (𝐶𝑃𝑓𝑐𝑡- − 𝐶𝑃𝑜𝑏𝑠-)I3-JK , (3)

where 𝐶𝑃𝑓𝑐𝑡- (𝐶𝑃𝑜𝑏𝑠-) represents the cumulative probability of the forecast (observations)

through category 𝑖, and 𝑛 represents the total number of categories. The RPSS is then (equation

4):

𝑅𝑃𝑆𝑆 = L1 − MNOPQRSTUVWMNOTXYZUWQXQ[\

] 𝑥100, (4)

where 𝑅𝑃𝑆 _`abcB. is the RPS for the forecast and 𝑅𝑃𝑆bd-ec._d_fg is the RPS for climatology. An

RPSS of 100% indicates perfect forecast skill, with positive (negative) values indicating the

forecast is more (less) skillful than climatology. The median RPSS value across all hindcast years

is reported.

Downscaling Global Dynamic and Regional Empirical Models

Dynamic Models

Numerous approaches for dynamically and statistically downscaling dynamical model output are

available, but not reviewed here. For this study, a quantile mapping approach to remove quantile

dependent biases based on cumulative distribution functions of predicted and observed

precipitation is adopted using drop-one-year cross-validation (Maraun 2013). Other approaches

for downscaling and bias-correction exist, such as linear interpolation, spatial disaggregation, or

downscaling with regional climate models, yet quantile mapping is selected here due to its

simplicity and low computational requirements (Zhao et al. 2017). Downscaling of NMME

precipitation predictions using model output statistics was also explored, yielding similar results.

Statistically corrected ensemble NMME JJAS precipitation predictions issued June 1 for Koga

(Figure 4, Table 1) illustrate relatively poor skill, indicating that the dynamic models are not well

13

capturing variability in seasonal precipitation at local scales. Additionally, predictions in some

years miss by two categories (an “extreme miss”; Table 1). Although quantile mapping, or similar,

is frequently applied bias-correction and post-processing (Bogner et al 2016), some argue that

these methods are incapable of bridging the scale mismatch (Maraun 2013) as inconsistencies aside

from model errors are introduced due to the inability to properly capture small-scale variability.

Inadequate representation of smaller scale variabilities could account for the lack of skill in the

bias-corrected NMME local predictions here and may warrant exploration of other bias-correction

techniques.

Figure 4. Bias-corrected JJAS NMME ensemble averaged precipitation predictions (blue

line), NMME ensembles (n = 99; boxes) and observations from CHIRPS (black line) for the

Koga (a, correlation = 0.21) and Finchaa (b, correlation = 0.41) regions, 1982-2017.

a

b

14

Table 1. Deterministic and categorical verification metrics for bias-corrected NMME

dynamic (1982-2017) and regional empirical (1982-2011) model JJAS precipitation

predictions, issued June 1 for the Koga and Finchaa regions.

Model Local

Region

Pearson

Correlation

Rank

Probability

Skill Score (%)

HIT Score, %

(# years)

Extreme Miss

Score, %

(# years)

Downscaled

Dynamic

(NMME)

Koga 0.21 -4.4 22.2 (8) 16.7 (6)

Finchaa 0.41 28 38.9 (14) 8.33 (3)

Empirical

Cluster

Regression

Koga 0.33 9.5 37.9 (11) 24.1 (7)

Finchaa 0.53 -11 41.4 (12) 6.9 (2)

Empirical Cluster Model Zhang et al. (2017) propose an empirical seasonal prediction model for the BNB using a k-means

clustering technique to assign homogeneous precipitation regions in an effort to decrease the larger

regional prediction scale adopted by the NMA of Ethiopia (Korecha and Sorteberg 2013; Zhang

et al. 2017). Once clusters are defined, the cluster level prediction is regressed to local regions

within the cluster using a principle component regression model (PCR) framework (Zhang et al.

2017). Local predictions of JJAS seasonal precipitation issued on June 1 demonstrate increased

correlation skill over NMME forecasts (Figure 5, Table 1) yet remain modest. Prediction

correlation skill varies across clusters and is particularly poor along cluster boundaries (Korecha

and Sorteberg 2013; Zhang et al. 2017). With the possible exception of Hit Score favoring the

empirical cluster model approach, the remaining performance metrics do not show marked

difference with NMME performance (Table 1). This moderate skill performance warrants

investigation of alternative approaches, particularly those that are locally-tailored.

15

Figure 5. Bias-corrected JJAS empirical precipitation predictions (blue line) and

observations from CHIRPS (black line) for the Koga (a, correlation = 0.33) and Finchaa (b,

correlation = 0.53) regions, 1983-2011.

Local Statistical Prediction Model

Model Framework

A number of climate variables potentially modulating local inter-annual precipitation and

streamflow through regional and global teleconnections are identified and evaluated as possible

predictors of seasonal precipitation/streamflow by correlating gridded climate variables with

observed JJAS precipitation (CHIRPS) or historical streamflow record at the case study location

(e.g. Figure 6). Potential predictors include sea surface temperature (SST), sea level pressure

(SLP), geopotential height (GH), meridional and zonal wind (MW), precipitable water content

(PWC), and surface air temperature (SAT). Climate variable regions with a mean leave-one-year-

out cross-validated correlation with observed JJAS precipitation at the 90% statistically significant

level are retained for possible inclusion into the prediction model. Various lead times (forecast

issue dates) are assessed (January 1, April 1, and June 1) for the precipitation predictions, using

a

b

16

only climate variable states prior to that date to better understand how prediction skill may change

with longer lead time (Table 2). In addition to these climate variables, past studies indicate that

temporal SST gradients, or tendencies, may also boost predictive skill (Diro et al. 2011a; Funk et

al. 2014). To investigate, each statistically significant SST region is spatially averaged and all

combinations of temporal SST tendencies from October – May are correlated with observed JJAS

precipitation. Spatial SLP tendencies representing the North-Atlantic Oscillation and Indian

Monsoon are also explored. Model predictors retained for each forecast issue date are presented in

Table 2.

Figure 6. Correlation between Koga JJAS total precipitation from CHIRPS and global May

sea-surface temperatures. Red boxes represent regions in the Indian Ocean (left) and

Tropical Pacific Ocean (right) showing higher correlations.

17

Principle Component Regression (PCR), a common statistical method to reduce dimensionality

and multicollinearity among large sets of data (Delorit et al. 2017; Lins 1985; Zhang et al. 2016,

2017), is adopted for the prediction model framework. PCR first transforms a set of variables

(predictors) into uncorrelated, orthogonal principle components (PCs), ordered in terms of the

amount of variance explained in the dataset. PCs explaining more than 10% of the overall dataset

variance are retained, following Kaiser’s rule as in other studies (Delorit et al. 2017; Kaiser 1960;

Zwick and Velicer 1986). In the second step, multiple-linear regression is performed using all

retained PCs as predictors, as in equation 5:

𝑦.i = 𝛽. + 𝛽.𝑃𝐶1. + 𝛽.𝑃𝐶2. + ⋯+ 𝛽.𝑃𝐶.𝑓𝑜𝑟𝑡 = 1𝑡𝑜𝑛, (5)

𝑒. = 𝑦.i − 𝑦., (6)

18

where, 𝑃𝐶. represents the selected PC at a given time step, 𝛽 values are the fitted regression model

coefficients, 𝑦.i is the predicted value of JJAS precipitation/streamflow annually, and 𝑒. represents

the residuals (predicted minus observed 𝑦. values) for each time step (t) from equation (5). To

evaluate historical performance (had the prediction model been in place), a hindcast is performed

in which, for each year 𝑡 = 1𝑡𝑜𝑛, the model is fitted in a leave-one-year-out cross-validation

mode, providing a single (deterministic) JJAS precipitation/streamflow prediction for each

observed year (Block and Rajagopalan 2007). Model residuals, 𝑒., from all hindcast years are

fitted to a normal distribution with mean zero, using leave-one-year-out cross-validation, and

added to the deterministic prediction values to form a distribution of predicted JJAS

precipitation/streamflow for each year (Delorit et al. 2017; Helsel and Hirsch 1992; Zhang et al.

2016, 2017). The reliability of the probabilistic ensemble was verified using reliability diagrams

(not shown; Hsu and Murphy 1986).

Using this PCR approach, JJAS precipitation predictions are developed for the forecast issue dates

of interest (June 1st, April 1st, and January 1st; and JJAS streamflow predictions for June 1st). Model

combinations for each lead time are compared using the Generalized Cross Validation (GCV)

score, balancing model error and number of predictors by rewarding reduction in prediction error

and penalizing overfitting (Block and Rajagopalan 2007; Craven and Wahba 1979). GCV is used

in model selection by choosing a subset of statistically significant climate variable predictors,

applying PCR, and then calculating the GCV using equation (7) below:

𝐺𝐶𝑉 =∑ SWp

11Wqr

(KsZ1)p , (7)

where 𝑒. is the residual,𝑁 is the number of observations, and m is the number of predictor variables

in the model. The model with the lowest GCV value for each lead time is retained.

Seasonal forecasts provide important indication of inflow volume during the dominant rainy

season, yet monthly streamflow forecasts may also be informative for water resource management

decisions. Following the PCR framework, the JJAS streamflow forecast is disaggregated to

monthly streamflow predictions using an analog approach. Given the JJAS prediction, a year with

19

the closest total JJAS streamflow is chosen from the historical record and the monthly distribution

of inflow is calculated. The predicted year is disaggregated to monthly predictions using the

distribution from the analog year. Since month-to-month streamflow persistence is relatively

strong following the JJAS season, monthly predictions for October, November, and December

were obtained by persisting the predicted inflow value forward. Inflow during January – May is

relatively consistent across the historical record, thus climatology is assumed for these monthly

streamflow values. Methods result in JJAS seasonal prediction and a monthly streamflow

prediction across all years.

The final set of predictors (Table 2) for the various leads for each region confirms many climate

teleconnections that influence moisture transport to the region, further warranting their inclusion.

For example, Pacific Ocean SSTs representing ENSO have been shown to influence JJAS seasonal

precipitation in Ethiopia (Camberlin 1997; Diro et al. 2011b; a; Elagib and Elhag 2011; Segele

and Lamb 2005). SSTs from the Indian and Atlantic Ocean basins also affect regional (Sahelian)

precipitation on inter-decadal timescales (Folland et al. 1986; Giannini et al. 2008; Hoerling et al.

2006; Lu and Delworth 2005). Regional pressure systems such as the St. Helena, Mascarene, and

Azores Highs have all been linked to JJAS moisture in Ethiopia (Black et al. 2003; Goddard and

Graham 1999; Latif et al. 1999; Segele and Lamb 2005; Shanko and Camberlin 1998; Viste and

Sorteberg 2013). Variations in Ethiopian climate have been strongly linked to the migration of the

ITCZ and pressure changes resulting from the Indian Monsoon, which can be captured through

SLP and local atmospheric variables such as PWC and SAT (Viste and Sorteberg 2013). Further,

migration of the ITCZ results in an inverse pressure relationship between the summer and winter

Indian Monsoon (Yancheva et al. 2007), thus a spatial wintertime SLP index across the Indian

Ocean may serve as a precursor to the expected strength of the summertime monsoon during the

JJAS season. In general, SST persistence is useful for inclusion in the model at most lead times,

and addition of more transient SLP and atmospheric variables provides valuable predictive

information at lead times closer to the season of interest.

Model Evaluation and Comparison

As expected, performance metrics typically decrease with increased lead time prior to the start of

the season of interest (Table 3). For example, cross-validated predictions for Finchaa precipitation

20

(Figure 7) correlate with CHIRPS at 0.59 for a forecast issued on June 1 (start of season), while

correlations decrease to 0.54 and 0.33, respectively, for April 1 and January 1 forecast issue dates.

Although longer-lead predictions show a decrease in skill, predictive information provided as early

as January may inform agricultural decisions that are made early in the calendar year (e.g. 22 years,

59% of the record, January and June predict the same category and 11 of those years categorical

prediction aligns with observed). Finchaa streamflow predictions correlate with observed

streamflow at 0.50 for a forecast issued June 1st (Figure 8). While a forecast issued June 1

incorporates SST, SLP, and atmospheric climate variables as predictors, more transient

atmospheric and SLP variables not surprisingly illustrate less skill (and are not included) for the

April and January forecasts given the longer lead times prior to the season of interest. Notably,

RPSS is positive for all lead times, indicating an improvement over climatology.

For Koga, a June 1 prediction uses the moderately developed, although not yet fully known, ENSO

state and includes SST, SLP, and atmospheric variables as predictors, resulting in a cross-validated

correlation of 0.57. Correlations drop substantially, however, for the April 1 forecast (0.13), likely

attributable to the well-documented “spring barrier” during which ENSO resets for the following

seasons (Chen et al. 2017; Clarke and Van Gorder 1999; Jin et al. 2008; Webster and Yang 1992).

Model skill at a January 1 forecast lead relies on SST climate signals from the northern hemisphere

fall and winter, which prove skillful for years when these signals persist into the following summer;

cross-validated predictions (Figure 7) correlate with CHIRPS at 0.52. RPSS values for all lead

times are positive. In 20 years, 54% of the record, January and June predict the same category

(aligning with the observed category 14 years), illustrating the possible value of longer-lead

predictions to inform crop decisions made early in the year.

21

Figure 7. Local statistical model JJAS precipitation predictions (median = blue line,

ensemble distribution = boxes) and JJAS precipitation from CHIRPS (black line) for the

Koga (a – June forecast date, b – April, c – January) and Finchaa (d – June, e – April, and f

– January) regions.

Figure 8. Local statistical model JJAS streamflow predictions (median = blue line, ensemble

distribution = boxes) and observed JJAS streamflow (black line) for the Finchaa (a – June

forecast date) and Amerti (b – June forecast data) Rivers.

22

For comparison, correlations of NMME predictions and CHIRPS precipitation for Koga (0.21,

0.15, and 0.02 for forecast issue dates of June 1, April 1, and January 1, respectively) and Finchaa

(0.41, 0.04, and 0.02 for June 1, April 1, and January 1, respectively) are notably less for all lead

times. These cumulative results suggest the superior performance of the local statistical model in

predicting local scale variability in precipitation.

Comparison of NMME and local statistical predictions at additional spatial and temporal scales

further illustrates the intricate link between model structure and prediction skill. Temporally, both

downscaled NMME and the local statistical model are less skillful in predicting a single month of

23

the JJAS season than in predicting the JJAS seasonal total precipitation, based on deterministic

Pearson correlation and categorical (using the full prediction ensemble) RPSS metrics (Figure 9a).

This increasing skill with temporal aggregation is not unexpected and has been posited by others

(e.g. Bohn et al. 2010). Perhaps more interestingly, performance metrics indicate that the local

statistical model is superior across all temporal prediction scales (Figure 9a).

Figure 9. Comparison of correlation (left y-axis, black) and RPSS (right y-axis, blue)

prediction skill for Koga from the NMME (o) and the local statistical model (*) for temporal

(a; 1, 2, and 4-month) and spatial (b; small – Koga basin, 1,110 km2, medium – MBNR, 49,284

km2, and large – BNB, 308,025 km2) prediction periods. All forecasts issued June 1.

Considering JJAS precipitation predictions at different spatial scales, the local statistical model is

superior at the small scale, while at the large scale NMME appears more suitable (Figure 9b).

However, NMME skill may increase at the medium and small scale with alternative downscaling

a

b

24

methods and statistical model prediction skill may increase if model structures are tailored to the

respective regions.

Based on this comparison, at a regional or basin scale, NMME prediction models may suffice,

however, at the local level, a tailored statistical prediction framework may be warranted. Skillful

prediction, however, does not necessarily directly translate into improved sectoral prediction and

decision-making.

25

CHAPTER 3: COUPLING PREDICTIONS WITH RESERVOIR MODELS

Application of Precipitation Predictions to Reservoir Management

Framework for coupling seasonal precipitation and reservoir volume

Coupling skillful precipitation predictions at the local scale with sectoral operations and

management may enhance local economic and environmental benefits. For example, a reservoir at

Koga allows for irrigated agriculture during a second cropping season (October – May), yet, given

interannual variability in precipitation, reservoir volumes at the end of the rainy season vary year

to year, which affects agricultural land preparation, seed procurement, planned water allocation,

etc. These decisions are made in advance of the end of the rainy season (October), thus the June

1 precipitation predictions and reservoir state are used to predict October reservoir volume.

General reservoir characteristics, monthly water release data, and reservoir stage data were

compiled from multiple sources and provided by project partners (Birhanu et al. 2014; Mott

MacDonald 2006). Following previous research in East Africa, monthly potential

evapotranspiration was estimated using radiation and temperature through the Hargreaves method

(Block and Goddard 2012; Droogers and Allen 2002; George H. Hargreaves and Zohrab A.

Samani 1985; Hargreaves and Samani 1982). The following reservoir water balance (equation 8)

was developed for simulation:

𝑉(𝑡) = 𝑉(𝑡 − 1) + 𝑃(𝑡) ∗ 𝑆𝐴(𝑡) − 𝐸𝑇(𝑡) ∗ 𝑆𝐴(𝑡) + 𝐼(𝑡) − 𝑅(𝑡), (8)

where for each monthly time step (t), V is reservoir volume (m3), P is total precipitation (m), ET

is evapotranspiration (m), SA is reservoir surface area (m2), I is inflow (m3), and R represents water

released from the reservoir (m3). Minimal information of initial hydrological conditions is

available for the Koga basin, most notably soil moisture conditions and groundwater interactions,

and are therefore not directly modeled as individual terms, similar to previous studies (Birhanu et

al. 2014; Mott MacDonald 2006; Reynolds 2013). However, inflow values effectively represent

direct flow from the Koga River as well as other fluxes (groundwater, soil moisture, infiltration,

etc.) not explicitly accounted for in this equation due to data limitations. Initial hydrologic

conditions may be less critical for long-lead predictions (greater than one month) and small-sized

basins, thus implicit inclusion of these terms with ‘inflow’ is assumed for this application (Li et

26

al. 2009; Shukla and Lettenmaier 2011). Observations of reservoir volume, precipitation, and

releases, along with estimated evapotranspiration, were used to determine inflow to the reservoir

during the JJAS season for each year of operation (2013-2017). Based on this, a direct relationship

between the total JJAS season precipitation-evapotranspiration difference and volume of inflow to

the reservoir during JJAS was also developed (equation 9):

𝐼yy7O = 𝐶K ∗ (𝑃yy7O − 𝐸𝑇yy7O) −𝐶I, (9)

where 𝐼yy7Ois the volume of inflow during JJAS and 𝑃yy7O − 𝐸𝑇yy7O represents the total JJAS

precipitation less the total evapotranspiration volume during the JJAS season. Coefficients 𝐶K and

𝐶I have mean values of 0.14 (0.08 – 0.21) and 60 (0.83 – 120) based on a drop-one-year cross-

validation fit of the available observed record (mean coefficients are used for hindcast). Although

this does not explicitly account for base flow, inflow from the Koga River is dominated by rainfall-

runoff processes during this season. October 1 reservoir volumes are then the result of observed

May reservoir volumes plus 𝐼yy7Ovalues. Although verification is limited to the years 2013-2017,

simulated October reservoir volumes do match observed reservoir volumes in terms of full versus

not full (reservoir filled in all years except 2015; Figure 10). Using CHIRPS data for 1981-2017

this relationship is extended to simulate the expected volume of JJAS inflow for historical years

and October 1 reservoir volumes, contingent on average May storage volumes from 2013-2016

(Figure 10; black line). This hindcast simulation suggests that had the reservoir been in place for

the period 1981-2017, it would have filled in approximately 62% of the years (greater than 50%

probability). Thus, an October reservoir state less than full is not unusual.

27

Figure 10. Hindcast (1981-2017) of simulated October Koga reservoir levels (black), mean

climatology (blue) and distribution of predicted October 1 reservoir volume (boxes) using

local statistical model precipitation predictions issued June 1. Full reservoir storage level

(83.1 million cubic meters) shown in red. Predicted probability of filling (%) shown above

each year (green = hit, red = miss).

Using the local statistical model to predict reservoir storage

Probabilistic local statistical model precipitation predictions issued on June 1 for the period 1981-

2017 are input into the framework to provide a hindcast of predicted October reservoir volumes.

JJAS inflow volumes (Equation 9), based on precipitation predictions and average

evapotranspiration values, were then added to an average May reservoir volume to illustrate how

the framework can provide a distribution of predicted October reservoir volumes (Figure 10).

Leave-one-year-out cross-validation is used for predictor selection, PCR model calibration, and

error distributions for the precipitation predictions; the inflow model is also constructed in leave-

one-year-out cross-validation mode to provide reservoir volume predictions. While not a robust

forecast informed reservoir simulation, the framework allows precipitation predictions to be

translated into information regarding reservoir stage. For irrigated agriculture communities that

rely on both precipitation and reservoir allocation, this simple translation coupled with the

precipitation prediction may provide a more complete indication of the likely amount of water to

be received during the coming season.

54%

91%

57%88% 92% 99%

78%

45%96%

92%

90%

93%

81% 79%

77% 86

%

91%

91% 97%

38% 51%

39% 54% 62%

75% 88%

86% 82%

37%56%

74%

72%

89%

95%

6%41%

51%

28

The predicted median October reservoir volume agrees categorically (fill or no fill) with the

reconstructed October reservoir volumes in 27 years (73% of the timeseries). The JJAS reservoir

state is correctly predicted (probability greater than 50%) on June 1 for four years in which

observational data is available (2013-2017) based on the median forecast (very close one year).

Percentage probability of filling for each year shows how ensembles can be used to determine a

prediction, where greater than 50% probability corresponds with a full reservoir (Figure 10). While

a strictly categorical fill or no fill prediction was used for this example, ensemble distributions

provide the opportunity to predict the approximate volume in reservoir for no fill scenarios. When

climatological precipitation is used as the forecast (naïve benchmark; average May storage used

for hindcast years, observed for operational years, to obtain climatological prediction), the

reservoir state is always predicted to be full (Figure 10). While this happens to be correct in four

of the five years, missing the 2015 drought may have serious consequences, particularly since the

reservoir is expected to not fill frequently (approximately 1/6 of all years; Figure 10).

In years that the reservoir would not have filled, according to simulated ‘observations,’ the mean

reservoir volume was 76.8 Mm3 (6.3 Mm3 short of full on average). For these same years, the

average predicted reservoir volume was 84.2 Mm3 (average deficit of 4.48 Mm3) based on the

median prediction. Further, the 2015 season was anomalously dry in the region and anecdotal

evidence states the reservoir was just over half full by the end of the JJAS season. Results indicate

October 2015 reservoir volumes of 57.0 Mm3 and 63.4 Mm3 for observed, median predictions.

The predicted reservoir volumes and deficits during years that the reservoir does not fill compare

remarkably well with the average volumes and deficits of the simulated observations. Full

validation and assessment of the framework performance are limited given the lack of long

observational records of inflow, operational decisions, and other data availability challenges.

Nonetheless, an indication of forecast value and performance is still strongly inferable (Zhao et al.

2017; Turner et al. 2017).

Coupling Streamflow Predictions with Operational Reservoir Models

Model of the Finchaa-Amerti Reservoir System

A model of the Finchaa-Amerti reservoir system was created to simulate movement of water

through the system, calculating hydropower generation and water allocation downstream. The

29

reservoir simulation uses historical inflow data, knowledge of current monthly release patterns,

precipitation, evapotranspiration, as well as known characteristics of the Finchaa-Amerti system

in an iterative water balance approach (equation 8). Minimal information regarding hydrologic

conditions (e.g. soil moisture and groundwater interactions) is available for the region, and thus

are omitted from the water balance calculation as in previous local studies in the BNB (Birhanu

2014; MacDonald 2006; Reynolds 2013).

Area, elevation, and volume relationships for both Finchaa and Amerti were generated based on

available data for use in the reservoir simulation (Eastern Nile Power Trade Program Study). The

Climate Hazards Group Infrared Precipitation with Stations (CHIRPS) gridded precipitation

product and estimation of evapotranspiration using the Hargreaves method are adopted, as in

Chapter two (Funk et al. 2015; Hargreaves and Samani 1982). Iterations of the reservoir model

were performed for calibration and to ensure that important constraints on the system (i.e. reservoir

volume, elevation and area, power generation) were met. The reservoir model was validated using

historical inflow records and current operations, comparison with results from a past study,

anecdotal information, and known power generation for some historical years (Belissa 2016;

Addisu 2018). The model of the reservoir system was used in the prediction-coupled operational

reservoir model, below.

Prediction-coupled Operational Reservoir Model The existing disaggregated statistical JJAS streamflow prediction and reservoir simulation model

are coupled to investigate the skill of a prediction-coupled operational reservoir model to aid water

resource decision-making under variable climate conditions. Four operational strategies are

explored: A) current operations, B) meeting agricultural demands, C) maximizing hydropower

production using local-scale streamflow predictions, and D) maximizing hydropower production

using local-scale streamflow predictions and the Amerti reservoir as additional storage (i.e.

multiple-reservoir). To couple the prediction and reservoir simulation models, an intermediate

optimization function is developed with the goal of determining optimal releases for the prediction-

coupled reservoir model at each timestep using single-objective linear optimization.

30

Forecasted monthly streamflow, precipitation and evapotranspiration climatology were input to

the reservoir system simulation framework. At each monthly timestep, the forecasted Finchaa

reservoir inflows for the following 12 months were fed to the optimization framework to determine

optimal release strategies. The optimal release value for the current month was then input to the

reservoir simulation to determine reservoir state (volume, surface area, elevation). Stepping

forward, reservoir state from the previous month is updated based on observed inflow values

before the next month is forecast and the process repeated. For example, in the beginning of May,

forecasted inflow for the next year is optimized to determine the May release. In June, May

reservoir state is updated with observed inflow information before using the 12-month forecast to

optimize and determine the release value for June (Figure 11). At all timesteps, environmental

river flows and water supply demanded downstream are required to be met. The model also

requires storage at the end time period to equal initial storage for each optimization step, and input

initial storage is assumed to be the cited normal operating level (536.4 Mm3; Belissa 2016). Amerti

is modeled primarily as a storage reservoir, only releasing water that would otherwise spill.

Figure 11. Model framework for the prediction-coupled reservoir model.

Multiple scenarios are input to the reservoir model to understand performance under varying

operational strategies. First, the current static operational release pattern is simulated as a baseline

Forecast Prediction-coupled reservoir model

PrecipitationEvapotranspiration

Update reservoir simulation with observed information

from previous timestep

Optimization using 12-month forecast

First release value

Reservoir simulation

Optimized release schedule for variable

climate conditions across the timeseries

Prediction-coupled Reservoir Model Framework

31

(Scenario A). Next, a static operational policy to meet agricultural demand is explored (Scenario

B). Finally, the local-scale disaggregated statistical streamflow forecast for Finchaa (developed in

Chapter two) is used in the prediction-coupled reservoir framework to investigate a dynamic

release schedule with the objective of optimizing release for maximum hydropower generation

using single-reservoir (Scenario C) and multi-reservoir (Scenario D) systems.

Measures for Assessment of Model Performance

Model performance is evaluated by identifying the degree to which simulated reservoir volumes

and release schedules meet the needs of reservoir users, with benefit defined in terms of energy

generation (hydropower) and volume of water allocated (agriculture). The range in monthly and

yearly hydropower generation is quantified, as well as the total energy production over the 37-year

timeseries. Further, performance during the most wet (1991) and dry (1983) years in the historical

record provides indication of model performance during extreme conditions. Criteria of reliability,

resilience and vulnerability have been used in multiple reservoir studies and shown to provide

helpful information on performance of the system being modeled (Jain and Bhunya 2008). Taken

together, these criteria are used to further assess model performance under varying scenarios.

Reliability is the probability of whether or not the system is in a satisfactory state, ranging from

negative one (unsatisfactory) to positive one (satisfactory), and shown by the annual reliability

metric (equation 11):

𝑅𝑒𝑙𝑖𝑎𝑏𝑖𝑙𝑖𝑡𝑦 = 1 − (|gac`}gac`

) (11)

where 𝐹𝑦𝑒𝑎𝑟 refers to the number of years of failure and 𝑇𝑦𝑒𝑎𝑟 refers to the total number of years.

The resilience criterion describes how quickly a system is likely to rebound from being in a failed

state. Ideally, a system would quickly respond from a state of failure to reach a satisfactory state

in a timely manner. The resilience of the reservoir system is calculated as the inverse of the mean

time that is spent in the unsatisfactory state, following Jain and Bhunya (2008), using equation 12:

32

𝑅𝑒𝑠𝑖𝑙𝑖𝑒𝑛𝑐𝑒 = �K8∑ 𝑑(𝑖)8-JK �

sK, (12)

where N refers to the number of failure events, and d represents the duration of failure event, 𝑖.

Vulnerability refers to the likely magnitude of a failure, should one occur, and is a measure of the

severity of damage during a failed event. For this work, vulnerability was assessed in terms of the

difference in water allocation (or energy production) demanded and provided, as in equations 13

and 14:

𝑉𝑢𝑙𝑛cf = 𝑣𝑜𝑙𝑢𝑚𝑒𝑑𝑒𝑚𝑎𝑛𝑑 − 𝑣𝑜𝑙𝑢𝑚𝑒𝑎𝑙𝑙𝑜𝑐𝑎𝑡𝑒𝑑, (13)

𝑉𝑢𝑙𝑛�g�`_ = 𝑒𝑛𝑒𝑟𝑔𝑦𝑑𝑒𝑚𝑎𝑛𝑑 − 𝑒𝑛𝑒𝑟𝑔𝑦𝑔𝑒𝑛𝑒𝑟𝑎𝑡𝑒𝑑, (14)

where 𝑉𝑢𝑙𝑛cf = vulnerability with respect to agriculture and 𝑉𝑢𝑙𝑛�g�`_ = vulnerability with

respect to hydropower. The water allocation demand for irrigated agriculture was defined as in

previous work (Belissa 2016) and the desired monthly energy was taken as the current energy

production during normal operations.

Evaluation of Prediction-coupled Reservoir Model Performance The coupled prediction-reservoir model illustrates benefits to competing users under different

operational Scenarios (A, B, C, D as introduced above). Modeling current operational strategies

(Scenario A) based on a static rule curve developed using decades-long records of hydropower

demand in the system results in moderate energy generation (22.0 – 65.4 GWh monthly, 21731

GWh total) and relatively low allocation for downstream users (6.55 – 40.0 Mm3 monthly and

9921 Mm3 total; Table 4). Performance of the Finchaa reservoir under current operational

strategies yields low reliability (0.38 average) and resilience (0.15 average) and high vulnerability

(6.68 average) criteria values, indicating that performance of the system with respect to both

hydropower and irrigated agriculture is moderate at best (Figure 12). As shown in Figure 13a,

there are many months in which benefits to hydropower are low and irrigation deficits high. While

the current operational policies may have met past management plans, Ethiopia is rapidly

33

expanding hydropower production through re-operation of existing facilities and newly built

infrastructure. Thus, a re-evaluation of production across the grid to determine improved

operational strategies moving forward is warranted.

Figure 12. Parallel line plot showing resilience, reliability, and vulnerability metrics for the

operational strategies investigated using the prediction-couple reservoir model framework.

Note: vulnerability metric results have been standardized, and metrics reflect the average

criteria value with respect to hydropower and agriculture.

34

Figure 13. Average monthly benefits to hydropower (dots) and water allocation deficits for

irrigated agriculture (bars) with lines denoting 25th and 75th percentile (lines) for a) current

operations, b) meeting agriculture demand, c) maximizing hydropower and d) maximizing

hydropower with a multiple-reservoir scenario.

Comparing Finchaa reservoir monthly releases based on current operations and agriculture

demands illustrates unmet allocation (Figure 14), on the order of 153 Mm3 annually. Thus,

Scenario B considers the performance of the Finchaa system re-operated to meet agriculture

demands as often as possible (where demands defined as in Belissa 2016). Interestingly, results

indicate greater benefit to both hydropower and agriculture under this scenario, as more water is

passed through the turbines before allocation for irrigated agriculture. Additional water passed

through the turbines, particularly in critical high flow months (September – November) results in

increased net benefits to hydropower in addition to irrigated agriculture in Scenario B (Figure 13b).

Average metrics across sectoral users display increased reliability (0.77; 49% improvement over

current operations, Scenario A), resilience (0.27; 56% improvement) and decreased vulnerability

(4.29; 64% improvement) of the system and show enhanced performance (Table 4). Total energy

production (24745 GWh) and water allocation for agriculture (12148 Mm3) are nearly 12.2% and

a b

c d

35

18.3% higher, respectively, than under current operational strategies, highlighting the benefit to

each.

Figure 14. Comparison of Finchaa releases based on current operations (black) and

agricultural demands (red; Belissa 2016); unmet demand shaded in grey.

Both Scenarios A and B utilize a static operational curve and do not consider expected upcoming

conditions. Given substantial inter-annual climate variability within the BNB (coefficient of

variation of rainy season precipitation is 9% near the Finchaa Reservoir), predicted inflow

conditions – leading to dynamic operations – may prove valuable, particularly in dry years and

those following.

To explore coupled prediction-reservoir operations (Scenario C), the disaggregated local-scale

statistical streamflow prediction introduced in Chapter two is integrated with the Finchaa reservoir

model with the objective of maximizing hydropower generation. A dynamic, or flexible,

operational strategy (i.e. flexible reservoir rule curve) and the integration of forecast information

serve to increase the performance of the reservoir system, as shown by moderately high reliability

and resilience metrics and low vulnerability (Figure 12, Table 4). Important to note, comparison

of scenario metrics in Figure 12 reflect the average with respect to hydropower and agriculture

users, thus limited information on hydropower demand reduces relative skill of the prediction

36

informed scenario and improved demand information may alter results. Scenario C yields higher

energy generation (range = 365 – 990 GWh annually; Table 4) and water allocation (range = 129

– 572 Mm3 annually; Table 4) than current operations (Scenario A; energy range = 237 – 624 GWh

and water allocation range = 40.0 – 303 Mm3 annually), indicating superior performance of

forecast-informed flexible operations. As Figure 13c illustrates, hydropower generation is

maximized during high flow months (July – December). A systematic increase in the minimum

yearly and monthly energy production and water allocation is critically important from users’

perspectives, as the value of improving upon unfavorable conditions often outweighs that of more

acceptable conditions (Tversky and Kahneman 1992).

Comparison of scenarios during climate extremes is also insightful for evaluating the benefit of

reservoir operational strategies for the end-user. For instance, under current operations (Scenario

A), energy production and water allocation are low in dry years and higher in wet years, as may

be expected (391 GWh and 146 Mm3 compared with 624 GWh and 303 Mm3; Table 5). Although

energy production and allocation are moderate during the dry years, the year directly following

shows greatly decreased production and allocation (e.g. 237 GWh and 40 Mm3 for 1984; Figure

15). A similar trend follows for Scenario B, yet additional allocation allows a slight increase in

post-dry year benefits.

Table 5. Prediction-coupled reservoir model benefits to the end-user, all scenarios during a a dry year (1983), post-dry year (1984) and wet year (1991). Model Scenario

Reservoirs Modeled Forecast Total Energy (GWh) Total Water Allocation

(Mm3) Dry Post-

Dry Wet Dry Post-Dry Wet

A. Current Operations

Finchaa -- 391 237 624 146 40 303

B. Meet Agriculture Demand

Finchaa -- 252 320 821 50.1 96.5 440

C. Hydropower Maximizing

Finchaa STAT 365 499 789 129 220 423

37

D. Hydropower Maximizing

Finchaa and Amerti

STAT 508 552 952 227 261 531

Note: AG - agriculture, HYD - hydropower, AVG - average of both users, STAT - statistical, seasonal forecast.

Figure 15. Annual energy generation (a) and water allocation (b) for the Finchaa reservoir

for current operations (Scenario A), meeting agricultural demand (Scenario B), and

maximizing hydropower (Scenario C) during dry (1983), post-dry (1984) and wet (1991)

years, respectively.

The benefits of Scenario C are particularly evident during extreme conditions. While energy

production and water allocation are at a minimum during very dry conditions, the ability to foresee

drought allows the operational policy to shift such that energy production and water allocation the

year following do not drop considerably (Table 5, Figure 15), as is the case in Scenario A. Granted,

in this case, the following year is a near-average year, however a similar outcome may also be

expected for two subsequent dry years, with the second year faring better when a flexible rule

based on predicted climate information is included. The challenge for Scenario A – conditioned

on a static rule curve – is that although it may perform well in the dry year in terms of energy

production and agriculture allocation, as significant volumes of water are passed through the

turbines, this results in a severe depletion of the reservoir and an inability to similarly provide in

the following year. Outcomes from Scenario C for dry years and the year following – in which the

a b

38

effects of the one dry year are spread across multiple years – is also seen as favorable by users who

typically desire minimizing inter-annual variability in energy production and agricultural

allocation. Comparing across Scenarios A, B, and C for a dry year, although benefits do not

necessarily always meet agricultural demands (shown by deficits in Figure 13c, d), Scenario C

performs best based on examination of total benefits.

Opportunity for Future Re-operation of the Finchaa – Amerti System

Scenario C utilizes all available water in the Finchaa reservoir in numerous years, and is water

limited, not energy production limited (turbine size). Thus, to expand energy production and

potentially better satisfy all downstream agriculture demands, additional supply is needed. In the

last couple decades, a reservoir was created on the adjacent Amerti River to serve as additional

storage for the Finchaa system. Scenario C is therefore again considered but for the multiple

reservoir (Finchaa – Amerti) system, named Scenario D.

Not surprisingly, the addition of the Amerti reservoir increases overall performance. High

reliability (0.64 average; 88% percent increase over Scenario C) and resilience (0.24 average; 75%

increase) as well as low vulnerability (2.30 average; 42% increase) illustrate the benefit of access

to additional water supply (Table 4). Energy generation (494 – 990 GWh yearly; Table 4) and

water allocation (221 – 556 Mm3 yearly) are the highest across all scenarios, particularly in regard

to minimum monthly and annual production/allocation, most often associated with dry years and

immediately following (Table 5; Figure 13; Figure 16). While Scenario D is clearly superior at

meeting sectoral demands in most years (Figure 16), there are years where Scenario B outperforms.

For example, Scenario D adopts releases in anticipation of drier years in 1994 – 1996, and thus

performs less well compared with Scenario B. However, given the static Scenario B releases full

agricultural demands, performance suffers in 1998 due to depleted reservoir volumes, with

performance still below that of Scenario D for the few years following. The Finchaa reservoir

alone does not have the inflow capacity to satisfy agriculture demands (26.3% of years demand is

met), and while the Finchaa-Amerti system performs better (63.2% of years demand is met), 100-

percent reliability is still not achievable considering the historical record (Figure 17).

39

Figure 16. Annual downstream water allocation under varying scenarios for the Finchaa

(solid) and Finchaa-Amerti (dashed) systems.

Figure 17. Annual inflow volume to the Finchaa, Amerti, and combined reservoirs. Annual

agricultural water demand denoted by red line.

40

CHAPTER 4: DISCUSSION AND CONCLUSIONS

The BNB in Ethiopia provides critical natural resources for the nation and continent, yet spatial

and temporal variabilities in climate adversely affect agriculture, energy, and other sectors.

Advanced predictions of seasonal moisture to the region during the dominant JJAS season may

provide information to enhance water resource decisions, yet the disconnect between the spatial

scale upon which skillful predictions are issued and the sectoral decision-making scale renders

current predictive information inadequate in many cases. Through development of skillful local-

scale statistical, seasonal precipitation and streamflow predictions, coupled with a multi-purpose

reservoir model, this work offers an alternative means to bridge the existing disconnect between

prediction and decision-making scales for benefit to agriculture and energy sectors in Ethiopia.

Chapter two focuses on evaluating local-scale precipitation predictions to potentially enhance

agricultural decision-making in a Blue Nile basin location using statistically downscaled NMME

dynamic models, a regional cluster-based model, and a local statistical model. Statistical

downscaling of dynamic models and the regional cluster-based model result in low prediction skill

at the local scale, as compared to the local statistical model predictions. While the use of dynamical

models is potentially appealing from an operational perspective, the modest skill and model

complexity are clear drawbacks. Statistical methods have been utilized for decades in seasonal

climate prediction, and while not a novel approach, results here are clearly superior, showing

promise for enhanced application to water resources management.

It is generally unsurprising that statistical methods conditioned on local scale teleconnections can

result in higher prediction skill at the local scale. One clear advantage of statistical approaches is

the ability to condition relationships of interest at the local scale using the historical record without

the need to fully capture complex climate dynamics at the global or regional scale. In a region with

many teleconnections to global climate phenomena, such as ENSO and the Indian Monsoon,

statistical methods may be better positioned to capture smaller-scale inter-annual factors that

modulate variability. For the BNB, Ethiopia, regional variability in precipitation and streamflow

can differ markedly from local variability. For example, correlation of Koga and Finchaa JJAS

precipitation with BNB region-wide precipitation are only 0.51 and 0.70, respectively. These

41

differences illustrate the importance of capturing heterogeneous local climate patterns that

diversely control moisture transport to relatively proximal locations. While global climate

predictors may be similar regionally, there may be important nuances. Statistical approaches have

demonstrated hydroclimatic prediction skill in a variety of settings (e.g. snow-dominated basins:

Delorit et al. 2017, drought-prone regions: Mortensen et al. 2018), suggesting transferability of

methods outlined here may be promising, however dominant, identifiable wet seasons (or high

flow seasons) and clear climate signals are still likely necessary.

It is important to note that alternative statistical or dynamic bias-correction approaches could be

evaluated, including linear interpolation, spatial disaggregation, or downscaling with regional

climate models, which may increase local scale prediction skill (Fowler et al. 2007; Maraun 2013;

Wood et al. 2004). However, this may only serve to increase complex resource requirements for a

limited return. Likewise, statistical modelling methods have limitations, such as short (37 years

in this case) time series of observational data that are potentially very sensitive to single anomalous

years. For example, to compare with drop-one-year cross-validation, the 37-year historical record

was split into a 19-year period for model calibration and 18-year period for model validation by

random selection (and the process repeated hundreds of times). For a June 1 forecast model, the

median correlation is 0.44 for the calibration period and only 0.07 for the validation period. This

drop in correlation value is not altogether surprising considering the limited number of years in

the time series but does highlight the potential sensitivity of models calibrated on a short

observational record. One avenue for future work is to explore the Community Earth System

Model (CESM; Kay et al. 2015; Ridge et al. 2013) with thousands of representative years to extend

the “observational” record and potentially better characterize modeling uncertainties.

Model selection is intricately linked with spatial and temporal prediction scale. Recognizing the

limitations of the current study, results indicate that both dynamic and statistical models have value

for hydroclimatic prediction, with the decision of model structure based on the scale at which

prediction information is required. Although many studies demonstrate value in predictions at the

local scale, few local-scale operational prediction models exist. This work shows promise for

bridging this gap through empirical forecast methods in Ethiopia.

42

One manner in which the value of hydroclimatic prediction models and information can be

assessed is by the degree to which predictions can aid water resource planning and management

decisions, particularly in preparing for extreme conditions. Multiple factors, including model

structure or complexity, prediction skill, scale, lead time, and uncertainty evaluated here do play

an important role in determining prediction value. As demonstrated, at larger (basin) spatial scales,

the NMME dynamic model structures outperform statistical prediction models based on both

deterministic and categorical metrics. Yet, for decisions at the local scale, statistical model

structures conditioned on oceanic and atmospheric climate variables better capture small-scale

variabilities in moisture transport to the region, exhibiting greater prediction skill for predictions

issued at the start of the season (June) and extending up to five months prior (January). Thus,

careful consideration of model structure based on the targeted decision scale has the potential to

increase the value of predictions to the end-user.

Further, Chapter three examines how local-scale hydroclimate predictions may be coupled with

reservoir models to translate predictive information into actionable strategies that may aid water

resource management decisions. Reservoir systems are a critical resource that support energy,

agriculture and other sectors, and may allow communities to mitigate risks posed by climate

variability. Yet, the management of reservoir systems is complex. In many cases, operational rule

curves have been developed to guide management of the system. Significant climate variability

can challenge static rule curve approaches. Two examples of integrating local-scale predictive

information with reservoir management models are evaluated in this work: 1) using JJAS

precipitation predictions to predict October irrigation reservoir volume, and 2) coupling local-scale

streamflow predictions with a multi-purpose reservoir model to evaluate alternative operational

strategies.

Precipitation predictions applied to a simple reservoir model for a June 1 prediction of end of

season (October 1) Koga reservoir volume would likely have correctly predicted whether or not

the reservoir would fill in 73% of years based on median predicted values, supporting farmers’

decisions regarding seed allocation and land preparation. Interaction with collaborators suggests

that early indication of October 1 reservoir volume, while simple, has great value. For example, in

2015 the region experienced one of the worst droughts in the historical record and the Koga

43

reservoir was half-full following the JJAS season. Although farmers were told to prepare only half

their land in expectation of decreased water allocation, this information was received too close to

the planting season to inform decisions; thus, huge land preparation and labor costs were lost.

Although skillful predictions do not necessitate enhanced decision-making on their own, they can

actively contribute to the suite of information available to help communities partially mitigate risks

associated with climate variability. Early indication in years where alternative decisions may be

warranted is imperative to allow for adaptive decision-making.

The coupled prediction-reservoir model of the Finchaa – Amerti system also exhibits skill in

maximizing energy production and meeting agricultural demands. The addition of new

hydropower facilities and agricultural expansion in the basin has warranted a re-evaluation of

system-wide operations, which this work begins to address. The value of flexible operational

strategies with predictive information is especially apparent in the dry years and following by

prescribing a reserved quantity of water in the dry year (more conservative allocation) to allow for

more available water in the subsequent year.

Chapter three examines a few alternative operational strategies, but clearly not an exhaustive list.

Expected energy and crop prices could inform additional optimization objectives and further

quantify the benefits and tradeoffs of different sectoral allocation quantities and timing and

operational strategies. Different energy production strategies, such as maximizing firm energy,

may also be identified. While the integrated models illustrate potential for increased hydropower

benefits on the Finchaa-Amerti system, this value can only be realized if there is a market for

additional power generation. Further, a more rigorous assessment of prediction uncertainty and its

translation through the reservoir model, may be warranted to better quantify the robustness of

results across different patterns of climate variability.

If the motivation for climate risk management is to support community adaption to climate

variability, then factors affecting forecast uptake – beyond forecast skill – must also be considered.

Social science and communication literature has repeatedly suggested that climate forecast

advancements are not readily integrated into local-scale decision making structures or utilized to

their full potential by the intended audiences (Gilles and Valdivia 2009; Pennesi 2007, 2011; Plotz

44

et al. 2017; Ziervogel et al. 2010), however this is not without its challenges, including prediction

uncertainty, scale mismatch between climate and prediction models, institutional or political

constraints, and risk aversion (Gilles and Valdivia 2009; Pennesi 2007; Ziervogel et al. 2010).

Communication and dissemination strategies for increased understanding and uptake suggest that

strategies tailored to the local region are key (Gilles and Valdivia 2009; Pennesi 2007, 2013;

Ziervogel et al. 2010). Thus, developing skillful forecasts at the local, decision-making scale, and

integrating with sectoral models, as demonstrated here, is imperative to resolve the disconnect

between the scale at which existing models are skillful and the scale at which prediction

information is valuable to end-users. This can serve to provide actionable information to aid water

resources planning and management decisions and reduce the vulnerability of communities to

climate variability.

45

REFERENCES Addisu, G. (2018). “Context and Background of the Koga, Finchaa, and Amerti Reservoirs.”

Personal Communication. Barnston, A. G. (1992). “Correspondence among the Correlation, RMSE, and Heidke Forecast

Verification Measures; Refinement of the Heidke Score.” Weather and Forecasting, 7, 699–709.

Bekele, F. (1997). “Ethiopian Use of ENSO Information in Its Seasonal Forecasts.” Internet Journal of African Studies, (2).

Belissa, A. (2016). “Establishing optimal reservoir operation of Fincha`a – Amerty Reservoirs.” Addis Ababa Institute of Technology.

Birhanu, K., Alamirew, T., Olumana Dinka, M., Ayalew, S., and Aklog, D. (2014). “Optimizing Reservoir Operation Policy Using Chance Constraint Nonlinear Programming for Koga Irrigation Dam, Ethiopia.” Water Resources Management, 28(14), 4957–4970.

Bisetegne, D., Ogallo, L., and Ininda, J. (1986). “Rainfall Characteristics in Ethiopia.” Proceedings of the First Technical Conference on Meteorological Research in Eastern and Southern Africa, UCAR, Nairobi, Kenya.

Black, E. (2005). “The relationship between Indian Ocean sea-surface temperature and East African rainfall.” Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 363(1826), 43–47.

Black, E., Slingo, J. M., and Sperber, K. R. (2003). “An observational study of the relationship between excessively strong short rains in coastal East Africa and Indian Ocean SST.” Monthly Weather Review, 74–94.

Block, P., and Goddard, L. (2012). “Statistical and Dynamical Climate Predictions to Guide Water Resources in Ethiopia.” Journal of Water Resources Planning and Management, 138(3), 287–298.

Block, P. J., Souza Filho, F. A., Sun, L., and Kwon, H.-H. (2009). “A Streamflow Forecasting Framework using Multiple Climate and Hydrological Models.” JAWRA Journal of the American Water Resources Association, 45(4), 828–843.

Block, P., and Rajagopalan, B. (2007). “Interannual Variability and Ensemble Forecast of Upper Blue Nile Basin Kiremt Season Precipitation.” Journal of Hydrometeorology, 8(3), 327–343.

Bohn, T. J., Sonessa, M. Y., and Lettenmaier, D. P. (2010). “Seasonal Hydrologic Forecasting: Do Multimodel Ensemble Averages Always Yield Improvements in Forecast Skill?” Journal of Hydrometeorology, 11(6), 1358–1372.

Camberlin, P. (1997). “Rainfall Anomalies in the Source Region of the Nile and Their Connection with the Indian Summer Monsoon.” Journal of Climate, 10, 1380–1392.

Camberlin, P., and Philippon, N. (2002). “The East African March – May Rainy Season : Associated Atmospheric Dynamics and Predictability over the 1968 – 97 Period.” Journal of Climate, 15(9), 1002–1019.

Chen, D., Zebiak, S. E., Busalacchi, A. J., and Cane, M. A. (2017). “An Improved Procedure for El Niño Forecasting : Implications for Predictability.” American Association for the Advancement of Science, 269(5231), 1699–1702.

Cheneka, B. R., Brienen, S., Fröhlich, K., Asharaf, S., and Früh, B. (2016). “Searching for an Added Value of Precipitation in Downscaled Seasonal Hindcasts over East Africa: COSMO-CLM Forced by MPI-ESM.” Advances in Meteorology, 2016, 1–17.

46

Clarke, A. J., and Van Gorder, S. (1999). “The connection between the boreal spring Southern Oscillation persistence barrier and biennial variability.” Journal of Climate, 12(2), 610–620.

Coelho, C. A. S., Pezzulli, S., Balmaseda, M., Doblas-Reyes, F. J., and Stephenson, D. B. (2004). “Forecast calibration and combination: A simple Bayesian approach for ENSO.” Journal of Climate, 17(7), 1504–1516.

Conway, D. (2000). “The Climate and Hydrology of the Upper Blue Nile River.” The Geographical Journal, 166(1), 49–62.

Conway, D. (2005). “From headwater tributaries to international river: Observing and adapting to climate variability and change in the Nile basin.” Global Environmental Change, 15(2), 99–114.

Craven, P., and Wahba, G. (1979). “Smoothing Noisy Data with Spline Functions: Estimating the Correct Degree of Smoothing by the Method of Generalized Cross-Validation.” Numerische Mathematik, 31, 377–403.

Crochemore, L., Ramos, M.-H., Pappenberger, F., and Perrin, C. (2016). “Seasonal streamflow forecasting by conditioning climatology with precipitation indices.” Hydrology and Earth System Sciences Discussions, 1–33.

Delorit, J., Gonzalez Ortuya, E. C., and Block, P. (2017). “A Framework for Advancing Streamflow and Water Allocation Forecasts in the Elqui Valley, Chile.” Hydrology and Earth System Sciences Discussions, 21, 1–30.

Diro, G. T., Grimes, D. I. F., and Black, E. (2011a). “Teleconnections between Ethiopian summer rainfall and sea surface temperature: Part II. Seasonal forecasting.” Climate Dynamics, 37(1), 121–131.

Diro, G. T., Grimes, D. I. F., and Black, E. (2011b). “Teleconnections between Ethiopian summer rainfall and sea surface temperature: Part I-observation and modelling.” Climate Dynamics, 37(1), 103–119.

Diro, G. T., Grimes, D. I. F., Black, E., O’Neill, A., and Pardo-Iguzquiza, E. (2009). “Evaluation of reanalysis rainfall estimates over Ethiopia.” International Journal of Climatology, 29, 67-78.

Diro, G. T., Tompkins, A. M., and Bi, X. (2012). “Dynamical downscaling of ECMWF Ensemble seasonal forecasts over East Africa with RegCM3.” Journal of Geophysical Research Atmospheres, 117(16), 1–20.

Droogers, P., and Allen, R. G. (2002). “Estimating reference evapotranspiration under inaccurate data conditions.” Irrigation and drainage systems, 16, 33–45.

Eklundh, L., and Pilesjö, P. (1990). “Regionalization and spatial estimation of ethiopian mean annual rainfall.” International Journal of Climatology, 10(5), 473–494.

Elagib, N. A., and Elhag, M. M. (2011). “Major climate indicators of ongoing drought in Sudan.” Journal of Hydrology, Elsevier B.V., 409(3–4), 612–625.

Eriksson, S. (2013). “Water quality in the Koga Irrigation Project, Ethiopia : A snapshot of general quality parameters.” Department of Earth Sciences, Air, Water, and Landscape Science, Uppsula University.

Faber, B. A., and Stedinger, J. R. (2001). “Reservoir optimization using sampling SDP with ensemble stream¯ow prediction (ESP) forecasts.” Journal of Hydrology, 249, 113–133.

Folland, C. K., Palmer, T. N., and Parker, D. E. (1986). “Sahel rainfall and worldwide sea temperatures, 1901–85.” Nature, 320(6063), 602–607.

47

Fowler, H. J., Blenkinshop, S., and Tebaldi, C. (2007). “Linking climate change modelling to impacts studies: recent advances in downscaling techniques for hydrological modelling.” International Journal of Climatology, 27, 1547–1578.

Funk, C., Hoell, A., Shukla, S., Bladé, I., Liebmann, B., Roberts, J. B., Robertson, F. R., and Husak, G. (2014). “Predicting East African spring droughts using Pacific and Indian Ocean sea surface temperature indices.” Hydrology and Earth System Sciences, 18(12), 4965–4978.

Funk, C., Peterson, P., Landsfeld, M., Pedreros, D., Verdin, J., Shukla, S., Husak, G., Rowland, J., Harrison, L., Hoell, A., and Michaelsen, J. (2015). “The climate hazards infrared precipitation with stations—a new environmental record for monitoring extremes.” Scientific Data, 2, 150066.

Gebrehiwot, S. G., Taye, A., and Bishop, K. (2010). Forest cover and stream flow in a headwater of the Blue Nile: Complementing observational data analysis with community perception. Ambio.

Gelati, E., Madsen, H., and Rosbjerg, D. (2014). “Reservoir operation using El Niño forecasts—case study of Daule Peripa and Baba, Ecuador.” Hydrological Sciences Journal, 59(8), 1559–1581.

Georgakakos, A. P. (1989). “The Value of Streamflow Forecasting in Reservoir Operation.” Water Resources Bulletin, 25(4), 789–800.

George H. Hargreaves, and Zohrab A. Samani. (1985). “Reference Crop Evapotranspiration from Temperature.” Applied Engineering in Agriculture, 1(2), 96–99.

Giannini, A., Biasutti, M., and Verstraete, M. M. (2008). “A climate model-based review of drought in the Sahel: Desertification, the re-greening and climate change.” Global and Planetary Change, Elsevier B.V., 64(3–4), 119–128.

Gilles, J. L., and Valdivia, C. (2009). “Local Forecast Communication In The Altiplano.” Bulletin of the American Meteorological Society, 90, 85–91.

Gissila, T., Black, E., Grimes, D. I. F., and 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., and Viste, E. (2017). “Seasonal predictability of Kiremt rainfall in coupled general circulation models.” Environmental Research Letters, 12(11), 114016.

Goddard, L., and Graham, N. E. (1999). “Importance of the Indian Ocean for simulating rainfall anomalies over eastern and southern Africa.” Journal of Geophysical Research, 104(D16), 19099.

Goddard, L., and Hoerling, M. P. (2006). Practices for seasonal-to-interannual climate prediction.

Goddard, L., Mason, S. J., Zebiak, S. E., Ropelewski, C. F., Basher, R., and Cane, M. A. (2001). “Current Approaches to Seasonal-to-Interannual Climate Predictions.” International Journal of Climatology, 21, 1111–1152.

Golembesky, K., Sankarasubramanian, A., and Devineni, N. (2009). “Improved Drought Management of Falls Lake Reservoir: Role of Multimodel Streamflow Forecasts in Setting up Restrictions.” Journal of Water Resources Planning and Management, 135(3), 188–197.

Gong, G., Wang, L., Condon, L., Shearman, A., and Lall, U. (2010). “A Simple Framework for Incorporating Seasonal Streamflow Forecasts into Existing Water Resource Management Practices.” JAWRA Journal of the American Water Resources Association, 46(3), 574–585.

48

Hamlet, A. F., Huppert, D., and Lettenmaier, D. P. (2002). “Economic Value of Long-Lead

Streamflow Forecasts for Columbia River Hydropower.” Journal of Water Resources Planning and Management, 128(2), 91–101.

Hargreaves, G. H., and Samani, Z. A. (1982). “Estimating Potential Evapotranspiration.” Journal of the Irrigation and Drainage Division, 108(3), 225–230.

Hastenrath, S., Polzin, D., and Camberlin, P. (2004). “Exploring the predictability of the ‘Short Rains’ at the Coast of East Africa.” International Journal of Climatology, 24(11), 1333–1343.

Helsel, D. R., and Hirsch, R. M. (1992). Statistical Methods in Water Resources. Elsevier Science Publishers B.V.

Hoerling, M., Hurrell, J., Eischeid, J., and Phillips, A. (2006). “Detection and attribution of twentieth-century northern and southern African rainfall change.” Journal of Climate, 19(16), 3989–4008.

Jain, S. K., and Bhunya, P. K. (2008). “Reliability, resilience and vulnerability of a multipurpose storage reservoir.” Hydrological Sciences Journal, 53(2), 434–447.

Jan van Oldenborgh, G., Balmaseda, M. A., Ferranti, L., Stockdale, T. N., and Anderson, D. L. T. (2005). “Did the ECMWF seasonal forecast model outperform statistical ENSO forecast models over the last 15 years?” Journal of Climate, 18(16), 3240–3249.

Jasperse, J., Ralph, F. M., Anderson, M., Brekke, L. D., Dillabough, M., Dettinger, M., Forbis, J., Haynes, A., Rutten, P., Talbot, C., and Webb, R. (2017). Preliminary Viability Assessment of Lake Mendocino Forecast Informed Reservoir Operations.

Jin, E. K., Kinter, J. L., Wang, B., Park, C. K., Kang, I. S., Kirtman, B. P., Kug, J. S., Kumar, A., Luo, J. J., Schemm, J., Shukla, J., and Yamagata, T. (2008). “Current status of ENSO prediction skill in coupled ocean-atmosphere models.” Climate Dynamics, 31(6), 647–664.

Jury, M. R. (2014). “Evaluation of coupled model forecasts of ethiopian highlands summer climate.” Advances in Meteorology, 2014.

Kaiser, H. F. (1960). “The Application of Electronic Computers to Factor Analysis.” Educational and Psychological Measurement, XX(1), 141–151.

Kalnay, E., Kanamitsu, M., Kistler, R., Collins, W., and Deaven, D. (1996). “The NCEP/NCAR 40-Year Reanalysis Project.” Bulletin of the American Meteorological Society, 77(3), 437–471.

Kay, J. E., Deser, C., Phillips, A., Mai, A., Hannay, C., Strand, G., Arblaster, J. M., Bates, S. C., Danabasoglu, G., Edwards, J., Holland, M., Kushner, P., Lamarque, J.-F., Lawrence, D., Lindsay, K., Middleton, A., Munoz, E., Neale, R., Oleson, K., Polvani, L., and Vertenstein, M. (2015). “The Community Earth System Model (CESM) Large Ensemble Project.” Bulletin of the American Meteorological Society, (August).

Kim, Y.-O., and Palmer, R. N. (1997). “Value of Seasonal Flow Forecasts in Bayesian Stochastic Programming.” Journal of Water Resources Planning and Management, 123(6), 327–335.

Kirtman, B. P., Min, D., Infanti, J. M., Kinter, J. L., Paolino, D. A., Zhang, Q., Van Den Dool, H., Saha, S., Mendez, M. P., Becker, E., Peng, P., Tripp, P., Huang, J., Dewitt, D. G., Tippett, M. K., Barnston, A. G., Li, S., Rosati, A., Schubert, S. D., Rienecker, M., Suarez, M., Li, Z. E., Marshak, J., Lim, Y. K., Tribbia, J., Pegion, K., Merryfield, W. J., Denis, B., and Wood, E. F. (2014). “The North American multimodel ensemble: Phase-1

49

seasonal-to-interannual prediction; phase-2 toward developing intraseasonal prediction.” Bulletin of the American Meteorological Society, 95(4), 585–601.

Korecha, D., and Barnston, A. G. (2007). “Predictability of June–September Rainfall in Ethiopia.” Monthly Weather Review, 135(2), 628–650.

Korecha, D., and Sorteberg, A. (2013). “Validation of operational seasonal rainfall forecast in Ethiopia.” Water Resources Research, 49(11), 7681–7697.

Lamontagne, J. R., and Stedinger, J. R. (2018). “Generating Synthetic Streamflow Forecasts with Specified Precision.” Journal of Water Resources Planning and Management, 144(4), 04018007.

Latif, M., Dommenget, D., Dima, M., Gr, Ouml, and Tzner, A. (1999). “The Role of Indian Ocean Sea Surface Temperature in Forcing East African Rainfall Anomalies during December & January 1997/98.” Journal of Climate, 12(12), 3497–3504.

Lins, H. F. (1985). “Interannual streamflow variability in the United States based on principal components.” Water Resources Research, 21(5), 691–701.

Lu, J., and Delworth, T. L. (2005). “Oceanic forcing of the late 20th century Sahel drought.” Geophysical Research Letters, 32(22), 1–5.

Lupo, A., and Kininmonth, W. (2013). “Global Climate Models and Their Limitations.” Climate Change Reconsidered II: Physical Science, 9–147.

MacDonald, M. (2006). Koga Dam and Irrigation Project Design Report Part 2: Irrigation and Drainage. Koga Irrigation Project, Ministry of Water Resources, Addis Ababa, Ethiopia, 137.

Maraun, D. (2013). “Bias Correction, Quantile Mapping, and Downscaling: Revisiting the Inflation Issue.” Journal of Climate, 26(4), 2137–2143.

Maurer, E. P., and Lettenmaier, D. P. (2004). “Potential Effects of Long-Lead Hydrologic Predictability on Missouri River Main-Stem Reservoirs*.” Journal of Climate, 17(1), 174–186.

Mott MacDonald. (2006). Koga Dam and Irrigation Project Design Report Part 2: Irrigation and Drainage.

Mwale, D., and Gan, T. Y. (2005). “Wavelet Analysis of Variability, Teleconnectivity, and Predictability of the September–November East African Rainfall.” Journal of Applied Meteorology, 44(1971), 256–269.

Nikulin, G., Asharaf, S., Magariño, M. E., Calmanti, S., Cardoso, R. M., Bhend, J., Fernández, J., Frías, M. D., Fröhlich, K., Früh, B., García, S. H., Manzanas, R., Gutiérrez, J. M., Hansson, U., Kolax, M., Liniger, M. A., Soares, P. M. M., Spirig, C., Tome, R., and Wyser, K. (2017). “Dynamical and statistical downscaling of a global seasonal hindcast in eastern Africa.” Climate Services, Elsevier, (November), 0–1.

Ntale, H. K., Gan, T. Y., and Mwale, D. (2003). “Prediction of East African seasonal rainfall using simplex canonical correlation analysis.” Journal of Climate, 16(12), 2105–2112.

Ogutu, G. E. O., Franssen, W. H. P., Supit, I., Omondi, P., and Hutjes, R. W. A. (2017). “Skill of ECMWF system-4 ensemble seasonal climate forecasts for East Africa.” International Journal of Climatology, 37(5), 2734–2756.

Pennesi, K. (2007). “Improving forecast communication: Linguistic and cultural considerations.” Bulletin of the American Meteorological Society, 88(7), 1033–1044.

Pennesi, K. (2011). “Making Forecasts Meaningful: Explanations of Problematic Predictions in Northeast Brazil.” Weather, Climate and Society, 3, 90–105.

50

Pennesi, K. (2013). “Predictions as Lies in Ceará, Brazil: The Intersection of Two Cultural Models.” Anthropological Quarterly, 86(20), 759–78943.

Philippon, N., Camberlin, P., and Fauchereau, N. (2002). “Empirical predictability study of October-December East African rainfall.” Quarterly Journal of the Royal Meteorological Society, 128(585 PART A), 2239–2256.

Plotz, R. D., Chambers, L. E., and Finn, C. K. (2017). “The Best of Both Worlds: A Decision-Making Framework for Combining Traditional and Contemporary Forecast Systems Climate and Oceans Support Program in the Pacific.” Journal of Applied Meteorology and Climatology, 56, 2377–2392.

Regonda, S. K., Rajagopalan, B., Clark, M., and Zagona, E. (2006). “A multimodel ensemble forecast framework: Application to spring seasonal flows in the Gunnison River Basin.” Water Resources Research, 42(9), 1–14.

Reynolds, B. (2013). “Variability and change in Koga reservoir volume, Blue Nile, Ethiopia.” Ridge, O., Hurrell, J. W., Alamos, L., and Berkeley, L. (2013). “The Community Earth System

Model: A Framework for Collaborative Research.” Bulletin of the American Meteorological Society, (September), 1339–1360.

Saha, S., Nadiga, S., Thiaw, C., Wang, J., Wang, W., Zhang, Q., Van den Dool, H. M., Pan, H. L., Moorthi, S., Behringer, D., Stokes, D., Peña, M., Lord, S., White, G., Ebisuzaki, W., Peng, P., and Xie, P. (2006). “The NCEP Climate Forecast System.” Journal of Climate, 19(15), 3483–3517.

Schepen, A., Wang, Q. J., and Robertson, D. E. (2012). “Combining the strengths of statistical and dynamical modeling approaches for forecasting Australian seasonal rainfall.” Journal of Geophysical Research Atmospheres, 117(20), 1–9.

Segele, Z. T., and Lamb, P. J. (2005). “Characterization and variability of Kiremt rainy season over Ethiopia.” Meteorology and Atmospheric Physics, 89(1–4), 153–180.

Segele, Z. T., Lamb, P. J., and Leslie, L. M. (2009). “Seasonal-to-Interannual Variability of Ethiopia/Horn of Africa Monsoon.” Journal of Climate, 22, 3396–3421.

Shanko, D., and Camberlin, P. (1998). “The effects of the southwest Indian ocean tropical cyclones on Ethiopian drought.” International Journal of Climatology, 18(12), 1373–1388.

Shukla, S., Funk, C., and 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).

Shukla, S., McNally, A., Husak, G., and Funk, C. (2014b). “A seasonal agricultural drought forecast system for food-insecure regions of East Africa.” Hydrology and Earth System Sciences, 18(10).

Shukla, S., Roberts, J., Hoell, A., Funk, C. C., Robertson, F., and Kirtman, B. (2016). “Assessing North American multimodel ensemble (NMME) seasonal forecast skill to assist in the early warning of anomalous hydrometeorological events over East Africa.” Climate Dynamics, 1–17.

Siam, M. S., and Eltahir, E. A. B. (2017). “Climate change enhances interannual variability of the Nile river flow.” Nature Climate Change, 7(5), 350–354.

Tefera, B., and Sterk, G. (2008). “Hydropower-Induced Land Use Change in Fincha’a Watershed, Western Ethiopia: Analysis and Impacts.” Mountain Research and Development, 28(1), 72–80.

51

Tversky, A., and Kahneman, D. (1992). “Advances in prospect theory: Cumulative representation of uncertainty.” Journal of Risk and Uncertainty, 5, 297–323.

Viste, E., and Sorteberg, A. (2013). “Moisture transport into the Ethiopian highlands.” International Journal of Climatology, 33(1), 249–263.

Webster, P. J., and Yang, S. (1992). Monsoon and Enso: Selectively Interactive Systems. Quarterly Journal of the Royal Meteorological Society.

Wood, A. W., Leung, L. R., Sridhar, V., and Lettenmaier, D. P. (2004). “Hydrologic implications of dynamical and statistical approaches to downscaling climate model outputs.” Climatic Change, 62, 189–216.

Yancheva, G., Nowaczyk, N. R., Mingram, J., Dulski, P., Schettler, G., Negendank, J. F. W., Liu, J., Sigman, D. M., Peterson, L. C., and Haug, G. H. (2007). “Influence of the intertropical convergence zone on the East Asian monsoon.” Nature, 445(7123), 74–77.

Yates, D. N., and Strzepek, K. M. (1998). “Modeling the Nile Basin under Climatic Change.” Journal of Hydrologic Engineering, 3(2), 98–108.

Yeshanew, A., and Jury, M. R. (2007). “North African climate variability. Part 3: Resource prediction.” Theoretical and Applied Climatology, 89(1–2), 51–62.

Zhang, Y., Moges, S., and Block, P. (2016). “Optimal Cluster Analysis for Objective Regionalization of Seasonal Precipitation in Regions of High Spatial–Temporal Variability: Application to Western Ethiopia.” Journal of Climate, 29(10), 3697–3717.

Zhang, Y., Moges, S., and Block, P. (2017). “Does objective cluster analysis serve as a useful precursor to seasonal precipitation prediction at local scale? Application to western Ethiopia.” Hydrology and Earth System Sciences Discussions, 1–18.

Ziervogel, G., Opere, A., Chagonda, I., Churi, J., Dieye, A., Houenou, B., Hounkponou, S., Kisiangani, E., Kituyi, E., Lukorito, C., Macharia, A., Mahoo, H., Majule, A., Mapfumo, P., Mtambanengwe, F., Mugabe, F., Ogalio, L., Ouma, G., Sall, A., and Wanda, G. (2010). Integrating meteorological and indigenous knowledge-based seasonal climate forecasts for the agricultural sector. Climate Change Adaptation in Africa Learning Paper Series, Ottawa, Canada.

Zwick, W. R., and Velicer, W. F. (1986). “Comparison of Five Rules for Determining the Number of Components to Retain.” Psychological Bulletin, 99, 432–442.