1. Introduction
Reservoirs are important water storage infrastructures that can effectively distribute water to serve the demand for hydroelectric energy, irrigation, municipal uses, ecosystem protection, and mitigating floods or droughts (Biemans et al. 2011; Bakken et al. 2016). Meeting such demands becomes more challenging with additional pressures generated by global trends in population growth and competing water demands for food, energy, and water (Brown et al. 2015; Rodell et al. 2018). Hence, maintaining sustainable management of water systems requires further flexibility of reservoir operations to support early actions and decisions. Reservoir operation is determined through a series of decisions over time, based on downstream demands and upstream flow (Labadie 2004). However, making such decisions is governed by the natural uncertainty in the reservoir inflow. Thus, using forecasts of future reservoir inflow can be advantageous to making efficient operating decisions (Ahmadi et al. 2015; Feng et al. 2017; Yang et al. 2020).
Hydrologic forecasts are typically generated via either dynamic, process-based climate models (e.g., Yuan et al. 2015) or via empirical, data-driven models (e.g., Block and Rajagopalan 2007). Both approaches have their limitations, with dynamic models often limited by resolutions and initialization procedures, while empirical models are constrained by short records and stationarity assumptions (Block and Goddard 2012). This study focused on the use of dynamical forecasts since they can take the underlying physical mechanisms into account, unlike statistical or data-driven models (Candogan Yossef et al. 2017; Woldemeskel et al. 2018). Uncertainties in dynamical forecasts can be reduced through a hybrid approach that benefits from statistical postprocessing (Wood and Schaake 2008). However, exploring such a hybrid approach is beyond the scope of our study. In dynamical forecasts, outputs produced by general circulation models (GCMs) are typically used as forcing for hydrological models. GCMs outputs are driven by large-scale climate drivers, such as El Niño–Southern Oscillation (ENSO) and sea surface temperature (SST). Advances in understanding such climate teleconnections have offered a significant improvement in predicting streamflow on a season ahead and at longer lead times (Hamlet and Lettenmaier 1999; Maurer and Lettenmaier 2004; Georgakakos and Graham 2008).
Skillful streamflow forecasts have proven to be useful for improving reservoir operations (e.g., Hamlet et al. 2002; Nayak et al. 2018; Giuliani et al. 2019; Ahmad and Hossain 2020). Using retrospective streamflow forecasts in operating reservoirs in the Columbia River, Hamlet et al. (2002) showed that long-lead streamflow forecasts can be effectively utilized to obtain increased annual average hydropower. During extreme conditions of droughts and floods, forecasting reservoir inflow informs dam operators how to efficiently manage the reservoir storage levels. While a drought event would require maximizing the storage in the dam (e.g., Golembesky et al. 2009), predicting a flooding event would inform the surplus storage to be released in order to accommodate incoming flow (e.g., Delaney et al. 2020). Furthermore, integrating seasonal climate forecasts with adaptive management schemes has demonstrated its potential for improving reservoir operation and provide adequate contingency measures during hydroclimatic disasters (Anghileri et al. 2016; Turner et al. 2017). Recently, the western United States had adopted a forecast informed reservoir operation approach to improve the reliability of reservoir operation by informing decisions about releasing or storing using forecasts (Jasperse et al. 2017).
The value of streamflow forecasting in reservoir operations becomes more significant in transboundary basins, where water availability regimes are impacted by water management practices in the upstream riparian countries (Munia et al. 2016; Shumilova et al. 2018). Installing an upstream dam in transboundary basins controls the inflow to the existing downstream dams. Thus, the operation of downstream dams is dependent on how a new upstream dam will operate during its filling and postfilling stages. One key question is how forecast value would change with introducing new dams upstream of an existing reservoir system in a transboundary basin. For example, how will the times of year when forecasting is deemed most important, i.e., higher forecast value, change during the filling and operation of a new upstream dam? Understanding such changes is important when incorporating forecasts in operation of an existing downstream dam.
A timely and compelling example is the upstream development in the Nile River basin (NRB) through the construction of the Grand Ethiopian Renaissance Dam (GERD). The current operation of reservoirs in the NRB is mostly based on rule curves that were designed from a time series of historical inflows collected during the pre-dam period. The efficiency of such rule curves is naturally limited as inflows substantially change with upstream projects during the post-dam period, e.g., the construction of the GERD. Therefore, informing the operation of downstream dams, e.g., the High Aswan Dam (HAD), with streamflow forecasts is one way to address such emerging challenges.
The value of streamflow forecasts can be undermined by limited information on operation of upstream projects. In such cases, when information on upstream dams is not made available or is still under negotiations, building scenarios of filling and operation of upstream dams can provide a realistic range of the potential value of streamflow forecasts in improving the operation of downstream existing dams. Consequently, this allows a tangible pathway for risk mitigation under uncertainty of inflow. This study proposes a framework, called Forecast-Based Adaptive Reservoir Operation (FARO), to integrate streamflow forecasts in deriving adaptive operating policy. The proposed framework is designed for existing reservoirs to operate in an adaptive manner that accounts for downstream demands while addressing upstream challenges from new transboundary dams. We used the HAD and GERD dams in the Nile River as examples of existing and planned dams, respectively, to test our FARO framework. This study addresses two primary questions:
What is the minimum forecast horizon for a reliable inflow forecast to be useful for HAD operations?
What is the potential value of forecast information at a long-term scale in adapting the HAD operation to future challenges introduced during filling and operation of GERD?
The rest of the paper is organized as follows. We present the study domain in section 2. We then describe the FARO framework and the experimental approach to assess the value of streamflow forecast for reservoir operations. In section 4, we discuss the results, highlighting the reservoir operation performance achieved by considering the contributions of monthly forecasts. We conclude the paper by summarizing the key findings and presenting further research directions.
2. Study area
The Nile River basin (NRB) is a major transboundary basin that extends across 11 countries in northeastern Africa (Fig. 1a). The NRB comprises two major tributaries, the Blue Nile (originates from Ethiopian plateau) and the White Nile (originates in the Great Lakes region of central Africa). The Blue Nile is the primary tributary to the main Nile River, providing 62% of the flow reaching Aswan (Zaroug et al. 2013), with three rainfall seasons; Bega, which refers to the dry winter season (October–February); Belg, the small rains of spring (March–May); and Kiremt, the wet summer season (June–September). The Kiremt season accounts for about three-quarters of the total annual rainfall (Tesemma et al. 2010; Taye and Willems 2012). The White Nile rises in the Great Lakes region of central Africa, with the most distant source in southern Rwanda, and flows north through Tanzania, Lake Victoria, Uganda, and southern Sudan. The Blue Nile and the White Nile meet at the Sudanese capital Khartoum and the Nile River then flows north through Sudan and Egypt to drain into the Mediterranean Sea.
(a) The Nile River basin with the location of HAD and GERD dams. (b) Climatological HAD and GERD inflow (averaged over 37 years 1981–2017) as modeled by the satellite-based framework developed by Eldardiry and Hossain (2019). The red and green lines represent the HAD target storage and demand used for deriving HAD optimal operation, respectively. (c) The annual GERD inflow anomalies averaged over the rainy season JJAS (June–September) during the period (1981–2017). The red dashed line represents the threshold for the dry vs wet year identification at half of the standard deviation of the JJAS total streamflow.
Citation: Journal of Hydrometeorology 22, 5; 10.1175/JHM-D-20-0241.1
The major dams on the Nile are Roseires Dam (Blue Nile in Sudan), Sennar Dam (Blue Nile in Sudan), High Aswan Dam (Main Nile in Egypt), and Owen Falls Dam (White Nile at Lake Victoria in Uganda). The High Aswan Dam (HAD) regulates the inflow to meet the downstream water supply for irrigation demands and hydropower generation. Lake Nasser (HAD reservoir) is one of the largest man-made lakes in the world and is vital to Egypt since it stores and regulates the Nile flow upstream of the HAD (Fig. 1a). The HAD reservoir has a total storage of 162 km3 with minimum and maximum operating levels of 147 and 182 m above mean sea level (MSL), respectively. At the beginning of the water year (1 August), the water level is kept at 175 m MSL (full supply level) to store incoming high flows (Moussa 2018). The storage increases gradually in the summer and, subsequently, the reservoir levels decrease from January to July as water is released (Fig. 1b).
The future hydropower dams inventory provided by Zarfl et al. (2015) reveals an increasingly hydropower development in the NRB. One of the ongoing projects is the GERD in Ethiopia that is currently under construction. The GERD construction started in 2011 with a plan to store water up to an elevation of 640 m (full supply level) corresponding to a storage capacity of 74 km3. The GERD will form the largest hydropower dam in Africa with a total capacity of 5150 MW supplied through 16 installed turbines, and it is likely to introduce a significant change in the electricity access for the entire Nile (Mulat and Moges 2014; Wheeler et al. 2018; Eldardiry and Hossain 2021). The GERD basin is the Upper Blue Nile basin located in the western part of Ethiopia (Fig. 1a), where most rainfall occurs during the summer months between June and September.
3. Methods and data
Figure 2 presents the main components of the FARO framework proposed in our study. First, FARO integrates meteorological forcing data from historical records and GCMs for driving a hydrological model. The hydrological model then produces streamflow for both historical and forecast periods. Second, we used the historical streamflow to define the benchmark streamflow forecasts: perfect forecast (assuming perfect knowledge of reservoir inflow) and streamflow climatology. The streamflow climatology is also used to correct the bias in the operational forecasts based on GCM forcings. Third, the streamflow forecasts are used with a model predictive control (MPC) scheme to optimize the reservoir operation. In MPC, we define different characteristics of the reservoir system, upstream conditions (e.g., introducing new dams), and downstream water system (e.g., demand scenarios). Finally, we assess the value of streamflow forecasting in informing adaptive reservoir operation. The different components of our FARO framework are described in detail in the next sections.
FARO framework for deriving adaptive reservoir operating policy using streamflow forecasts.
Citation: Journal of Hydrometeorology 22, 5; 10.1175/JHM-D-20-0241.1
a. Streamflow forecasting
1) Benchmark forecast
To understand the value of a forecast-based operating policy, we need to compare that performance against a benchmark. We applied a benchmark operating schemes using perfect streamflow forecasts. The perfect forecast is based on monthly reservoir inflow using historical simulated streamflow at HAD for the 10 years from 1993 to 2002. The HAD inflow is simulated using the Variable Infiltration Capacity (VIC) model [described later in section 3a(3)] and driven by Climate Hazards Group Infrared Precipitation with Station Data (CHIRPS) precipitation. Perfect forecast reflects the theoretical maximum benefit (or potential value) that could be achieved assuming we have full and perfect information on the future in the present when operational decisions are made. We also compare our results when using the climatology of streamflow, i.e., the average monthly streamflow simulated by the VIC model for 37 years from 1981 to 2017. The climatology infers the absence of information on future forecasts and thus represents a nonadaptive approach to develop the reservoir operational rules.
2) NMME forecast
We used the predicted monthly precipitation and temperature from North American Multi-Model Ensemble (NMME; Kirtman et al. 2014) to feed the hydrological model and produce streamflow forecasts for the 10-yr period from 1993 to 2002. The NMME real-time forecasts and hindcasts are available through the International Research Institute (IRI) for Climate and Society data portal. Archived forecast variables include precipitation, SST, and 2-m air temperature. A detailed list of experimental setup, available models, number of ensembles, and hindcast period can be found in Kirtman et al. (2014). To account for forecast uncertainty, we considered eight NMME models and each model includes between 4 and 24 ensemble members with maximum lead times of 9 (NASA-GEOSS2S), 10 (NCEP-CFSv2), and 12 (all other models) months. In total, multimodel ensemble utilized in this study consists of 90 members (see Table 1 for a summary of the eight models).
List of NMME models used in this study.
3) Hydrological modeling (VIC)
The predicted surface meteorological conditions (e.g., precipitation and temperature) are used as forcings for the VIC model to produce retrospective streamflow forecasts. VIC is a macroscale physically based, semidistributed hydrology model that closes the energy and water balances and includes subgrid variability in elevation, vegetation, and infiltration (Liang et al. 1994, 1996). The VIC setup used in our analysis was developed by Eldardiry and Hossain (2019) for the Blue Nile basin (BNB) at 0.1° (~10 km) spatial resolution using high spatial and temporal resolution of satellite observations, e.g., SRTM, CHIRPS, and MODIS. The forcing for historical and NMME forecast runs are both spatially downscaled to the grid scale of 0.1° using inverse distance weighted interpolation method. Eldardiry and Hossain (2019) have demonstrated the skill of the VIC model with satellite observations for simulating streamflow along the BNB [validated at Khartoum and Eldiem stations with a Nash–Sutcliff efficiency (NSE) of 0.68 and 0.92, respectively]. The reader is referred to Eldardiry and Hossain (2019) for further details on the VIC modeling framework over the BNB.
Because of the uncertainty in the hydrologic model and the uncorrected errors in downscaled forcings, a hydrologic postprocessing can produce more skillful and reliable hydrologic forecasts (Ye et al. 2014). We employed a lead-time-dependent (LTD) bias correction method to correct the errors in NMME streamflow forecasts (Jabbari and Bae 2020). The LTD bias correction is applied for the streamflow forecast using the ratio between the NMME-based streamflow (for 10-yr period 1993–2002) and the climatology of a reference streamflow (average streamflow for 37-yr period 1981–2017). In our study, a reference streamflow is produced using VIC simulations that are driven by CHIRPS precipitation for 37 years (1981–2017).
b. Model predictive control
To maximize the reservoir releases while incorporating inflow forecasts, we adopted a model predictive control formulation. MPC is a well-established optimal control technique that has recently gained increasing attention in the reservoir operation literature (Breckpot et al. 2013; Galelli et al. 2014). MPC is typically implemented in reservoir operations with a rolling horizon decision approach. This approach updates forecasts and decisions with each time step leading to more reliable operation (Zhao et al. 2012; Wan et al. 2016). The steps to derive optimal reservoir operation using MPC are as follows:
At each decision time instant t, a control problem for the reservoir operation (in our study, we used a water balance model controlled by target storage level and downstream demand) is formulated over a finite horizon from t to t + h, where h is the forecast horizon in months.
Applying a deterministic dynamic programming approach, we used a set of discrete finite horizon of water levels (or storage volumes) with 0.01-m increments to derive alternative sequences of release decisions.
An optimal release decision sequence (R1, R2, …, Rh) is then derived by minimizing the operation cost (i.e., the objective function of the optimization problem) over the forecast horizon h.
Only the release decision at the first time step R1 is used for the next time step (t + 1) (since a new forecast is issued).
A new optimization problem is then solved over the next horizon (t + 1: t + h + 1).
For the validation of the MPC model, we forced the model with observed time series of HAD inflows. The reason for using the observed data is to focus our validation only on the performance of the MPC formulation, i.e., the reservoir mass balance and optimization approach, without introducing uncertainties from data sources, e.g., using modeled streamflow or satellite-based water levels. Figure 3 shows the comparison of the observed HAD storage levels and outflow with the time series calculated using our MPC formulation during a 5-yr validation period (1998–2002). The comparison depicted a high skill modeling of HAD operation using the MPC formulation with an NSE and coefficient of determination (R2) of 0.80 and 0.94 for the HAD storage level, respectively (NSE = 0.95 and R2 = 0.95 for HAD outflow). The slight differences noticed in the storage levels are attributed to the water balance equation used to model the reservoir operation, which cannot perfectly mimic the actual mass balance. For instance, the increase in HAD storage is attributed to high inflow at the HAD reservoir in 1998 and 1999, that was considered as “high floods” for HAD operation (Sadek and Aziz 2005). These flooding events led to the formation of Toshka Lakes (located at the west of HAD reservoir) and the construction of spillway canals to link HAD reservoir with Toshka Lakes (Moussa 2018; Chipman 2019). Overall, the MPC is able to reasonably capture the actual reservoir operation with high skillful predictability of interannual and intra-annual storage and flow variations. The ability of the MPC model to simulate the water levels and releases at HAD is critical to further investigate the viability of the forecast information in adapting the HAD operation.
Validation of the MPC model for simulating storage water level and outflow of the High Aswan Dam during the five years (1998–2002).
Citation: Journal of Hydrometeorology 22, 5; 10.1175/JHM-D-20-0241.1
c. Upstream planned dams
One of the primary questions addressed in our FARO framework is to assess the value of forecast information to operate an existing dam when a newer dam is planned upstream. This value is more substantial during the filling phase of new dam when downstream countries expect a reduction in water supply as water is stored in an upstream reservoir. The HAD–GERD case in the Nile River basin typifies a timely example of transboundary challenge where introducing a new upstream dam would require revisiting the operation of existing dams. We used simulated streamflow at Eldiem station (i.e., GERD inflow) during the period 1981–2017 to build different scenarios of GERD filling (ranging from 2- to 12-yr filling scenarios). The GERD filling scenario follows an approach that assumes monthly filling of the dam in a pattern identical to its inflow, i.e., higher storage in summer months. We considered filling the GERD dam using a moving time window of N years for each N-yr filling scenario. For example, for a 3-yr filling scenario, we produced 35 possible hydrologic sequence to fill the dam in 3 years (the first window is between 1981 and 1983 and the last window is between 2015 and 2017). The GERD reservoir is considered completely filled when the reservoir storage reaches its maximum capacity of 74 km3. We here showed only the results when considering the average of the hydrologic sequences produced for 3- and 7-yr filling scenarios.
The GERD releases during the filling scenarios are routed downstream to Khartoum station and then added to the flow from the White Nile and Atbara River to form the total inflow into HAD. As future operating rules for the GERD are still under negotiation, we used a discrete dynamic programming approach to derive an operating curve for GERD that is based on optimal hydropower production. More details on the filling and operation scenarios can be found in Eldardiry and Hossain (2020).
d. Experimental setup
Our experimental approach considered three forecast alternatives: perfect forecast, streamflow climatology, and actual forecasts using NMME models (Fig. 2). Perfect forecast is used as our benchmark to assess the forecast value when using NMME- versus climatology-based forecasts. Our scenarios include also testing three demand trajectories: 25th, 50th, and 75th percentiles (calculated as percentiles of measured HAD outflow during the period 1993–2002) to represent low, normal, and high downstream demand patterns, respectively. Table 2 summarizes the scenarios included in our FARO experimental setup. The aim of such scenarios is to elucidate the impacts of reservoir operating behavior and upstream conditions on the usefulness of applying long-term forecasts.
Summary of the scenarios tested for different FARO inputs.
e. Evaluation of forecast value and skill
4. Results and discussion
a. NMME streamflow forecasts
The ability to make skillful hydrological forecasts is determined by different factors including hydrological initial conditions (mostly soil moisture) and the accuracy with which hydrological forcings are known (Lettenmaier 2017). Instead of using observed discharge, we here opted to use streamflow produced from VIC simulations using CHIRPS precipitation as our reference of evaluating the streamflow forecast. This comparison allows focusing our evaluation on the effects of forcing data, i.e., precipitation and temperature, without further including uncertainties associated with hydrologic modeling. Figure 4a shows the relative bias for the monthly streamflow at Khartoum station (the outlet of the Blue Nile) at different lead times ranging from 1 to 12 months. The NMME models behave differently with wide range of variations when compared to the reference streamflow. The NASA-GEOS-S2S model shows an underestimation in monthly streamflow for lead time greater than 2 months. The largest streamflow overestimation is noticed by the Canadian models: CanCM4i and CMC2-CanCM4 with an average relative bias (averaged over all the lead times) of 430% (16.86 km3 month−1) and 428% (16.34 km3 month−1), respectively. The skill performance of NMME forecasts is predominantly driven by strong seasonal- and lead-dependent biases (Slater et al. 2019).
Relative bias (%) of monthly streamflow (km3 month−1) at Khartoum station simulated in VIC using NMME forcings for the period (1993–2002). The NMME-based streamflow is compared to reference simulated streamflow using CHIRPS precipitation. The relative bias is shown (a) before and (b) after applying LTD bias correction.
Citation: Journal of Hydrometeorology 22, 5; 10.1175/JHM-D-20-0241.1
Figure 5 shows an example of 5-yr (1998–2002) bias-corrected streamflow for the eight NMME models at three lead times (1, 3, and 6 months). The bias correction approach resulted in high skillful streamflow with an average relative bias (averaged over both the lead time range and the eight models) of 1.3% (Fig. 4b). The GEM-NEMO model has the highest performance skill followed by the NCEP-CFSV2 with an average NSE of 0.89 and 0.88, respectively. The high skill of the NCEP-CFSV2 model has also been depicted over other regions, e.g., China (Ma et al. 2015) and the continental United States (Misra and Li 2014). When considering only the flooding season (June–September), we noticed an overall reduction in the NSE and correlation coefficient with an average of 0.54 and 0.79 (again averaged over the eight models and the lead times). A full range of the performance metrics (RBIAS, NSE, R2, and NRMSE) is provided in Tables S1 and S2 in the online supplemental material.
LTD bias corrected streamflow at Khartoum station using VIC simulations driven by NMME forecasts. Each time series is produced by concatenating the NMME-based streamflow at N lead time (where N is the lead time in months).
Citation: Journal of Hydrometeorology 22, 5; 10.1175/JHM-D-20-0241.1
b. Potential forecast value for HAD operation
We here used the perfect forecast benchmark to define the potential forecast value for HAD operation. Figure 6 shows the normalized penalty cost for optimal HAD operation under a normal demand trajectory (i.e., 50th percentile) and using perfect forecast at different lead times. For most months, except for flooding season in August and September, lower penalty costs are associated with integrating forecast information at longer lead time into our reservoir optimal operation. For instance, when using a perfect streamflow forecast in the months following the flooding season (October–December), the average normalized penalty cost ranges between 0.12% (1-month lead time) and 0.02% (12-month lead time). While the penalty cost decreases as streamflow forecast is available for longer lead time, this pattern is reversed for the HAD in August and September when the peak inflow reaches Lake Nasser. Such a contrasting pattern is attributed to the storage level objective that controls the optimization problem during the flooding season and penalizes the operation for being lower or higher than the target storage level. Thus, when predictions are available during the flooding season, the dam would store more water (i.e., requires higher storage levels exceeding the target storage) to offset the lower future inflows. The requirement of excess storage affirms the HAD being supported with Toshka spillway to the west of Lake Nasser where excess flow can be discharged.
Normalized penalty cost for HAD operation using perfect-based forecast (the potential value) averaged over the period 1993–2002 and at a demand trajectory of 50th percentile.
Citation: Journal of Hydrometeorology 22, 5; 10.1175/JHM-D-20-0241.1
We also considered high- and low-demand trajectories downstream of HAD, i.e., 25th and 75th quantiles of the HAD outflow to indicate potential levels of stress on the system downstream of HAD. Unsurprisingly, the penalty cost increases with higher-demand trajectory since more penalty will be induced to the supply objective function, i.e., less water to meet downstream demands. For instance, the normalized penalty cost increases from 0.06% to 0.17% when operating HAD with changing the demand trajectory from the 25th to 75th percentile and using perfect streamflow forecast at a lead time of 2 months.
Figure 7 shows the FVA to HAD operation (averaged over the nonflooding season from October through May) when using perfect streamflow forecast under different demand scenarios. The FVA has two main tipping points identifying the forecast lead time when FVA to HAD operation: 1) reaches its maximum value and 2) reaches zero, i.e., reservoir forecast horizon. For the demand scenarios at the 50th and 75th percentiles, the maximum FVA is achieved at a forecast of 4-month lead time (e.g., average FVA = 52% for the 75th percentile demand scenario). At lower than 4-month lead time, the FVA is less significant for higher demand scenarios (e.g., average FVA of 2-month lead time is 38.3% for 75th percentile demand scenario). This can be explained by the fact that downstream of HAD is subjected to high stress (due to high-demand scenario) and therefore adding forecast information at short lead time (up to 3 months) will not significantly impact the operation decision compared to forecasts at longer lead time. Conversely, in case of low-demand scenario where downstream system has less stresses, the HAD operation can significantly benefit from short lead time forecasts with high FVA (e.g., average FVA of 2-month lead time is 52.3% for the 25th percentile demand scenario). The FVA to HAD operation recedes gradually with longer lead-time forecasts up to a point when it reaches zero. Beyond this point, the reservoir operation cannot be further improved with forecasts at longer lead time (Fig. 7). The forecast horizon for HAD operation ranges between 5 and 12 months lead time for low- and high-demand scenarios. The long forecast horizon of HAD is attributed to its large storage capacity (average active storage–inflow ratio is equal to 0.95) that is sufficient to regulate the hydrologic variability of inflow and hedge for future demands.
Average FVA calculated as the reduction in month-over-month optimal cost using perfect forecast under different demand trajectories (considering only nonflooding seasons months from October through May).
Citation: Journal of Hydrometeorology 22, 5; 10.1175/JHM-D-20-0241.1
c. Skill of NMME-based forecasts
Figure 8 shows the range of the monthly penalty cost based on the actual forecasts from the eight NMME models at different lead times (only 1-, 3-, and 6-month lead times are shown) and compared to the perfect and climatology-based forecasts. The range of penalty cost from the NMME forecasts is able to capture the perfect forecast (blue line in Fig. 8). For instance, the normalized penalty cost when using NMME-based forecast in November with 3-month lead time ranges between 0.04% and 0.17%, which captures the potential value of the perfect forecast (0.07%). The differences between perfect and NMME forecasts are expected due to the uncertainty in NMME models that resulted in different forecast quality. Figure 9 shows the FSS [defined in Eq. (6)] as a function of the lead time (months) and the month during which the forecast is issued. NMME-based forecasts tend to overestimate the operation cost at longer lead times (comparing the average of the eight models with the perfect forecast). Higher skill of NMME forecasts is noticed in the winter months at shorter lead times, i.e., FSS is closer to zero. When using the climatology-based forecasts (green line in Fig. 8), a significant underestimation in the penalty cost is noticed as compared to the perfect forecasts (Fig. 9a). For example, the forecast skill in December when using NMME-based forecasts at 1- and 6-month lead time ranges between −0.22 km6 (FSS for climatology-based forecast is −11.1 km6) and 4.9 km6 (FSS for climatology-based forecast is −2.1 km6), respectively. The forecast skill does not decrease uniformly with lead time, and the selection of a skillful lead time varies as a function of forecast issuing month (flooding versus nonflooding seasons). Overall, when using actual forecasts from NMME models from October through December, HAD operation was improved with an average gain in forecast skill ranging from 16% (1-month lead time) to 231% (12-month lead time) as compared to the climatology-based forecasts.
Normalized penalty cost of HAD operation (averaged over the years 1993–2002) using perfect and NMME-, and climatology-based streamflow forecasts (only 1-, 3-, and 6-month lead times are shown).
Citation: Journal of Hydrometeorology 22, 5; 10.1175/JHM-D-20-0241.1
FSS when using (left) climatology- and (right) NMME-based forecasts for HAD operation as compared to perfect forecast benchmarking (note that the two panels are represented by different color scales).
Citation: Journal of Hydrometeorology 22, 5; 10.1175/JHM-D-20-0241.1
d. Examples for FARO application
1) Wet and dry years
To demonstrate the value of our FARO framework, we evaluated the forecast value when HAD operates in a dry year following a wet year. The dry and wet years are defined at Eldiem station (i.e., at the outlet of the Upper Blue Nile basin) by setting a threshold for the total discharge anomalies of the rainy season (June–September) for a 37-yr streamflow record between 1981 and 2017 (Fig. 1c). The threshold is set as equal to half of the discharge anomaly standard deviation: any discharge anomaly above 3.64 km3 month−1 is considered as a flooding condition (wet year), and any discharge anomaly below −3.64 km3 month−1 as a drought condition (dry year). In our example, we showed the forecast-based HAD operation for the two years 2001 and 2002. The two years represent upstream wet (2001) and dry (2002) conditions with a total flow volume at Eldiem station during the rainy season of 49.6 and 33.6 km3, respectively (Fig. 1c). The different climate is also evidenced by the storage water level observed by radar altimetry over lake Nasser (figure not shown). The water storage reaches its peak in October at a level of 180.59 and 177.98 m MSL in 2001 and 2002, respectively.
Following the FARO framework, the MPC model was forced by three streamflow forecast scenarios: perfect, climatology, and NMME-based forecasts, in the wet (2001) and dry (2002) years. As illustrated by Fig. 10, the streamflow forecasts at longer lead time can substantially ameliorate the dam operation in the dry year by storing more water and thereby reduce the operation penalty cost (when compared to observed operation). For instance, if considering forecasts at 3-month lead time, the total normalized penalty cost (summed for the two years 2001 and 2002), dropped down from 3.20% for the observed HAD operation to 0.91% in case of perfect forecast (and 0.83% for the average of eight NMME models). The reduction in the operation cost elucidates the opportunities to better operate the HAD reservoir using long-term streamflow forecasts to make decisions on storage and releases that would minimize the impact of dry years. For example, an adaptive operation is highly recommended when releasing water in the months preceding the flooding season, i.e., when the reservoir empties its storage. The HAD operation in 2002, for example, shows a significant drop in storage volume in August (119.10 km3), while forecast reservoir inflow suggests a higher storage level (132.76 km3) that can compensate for the inflow reduction in a dry year. While the MPC model with forecast information can potentially improve the HAD operation, slightly higher costs, are produced in the wet year (2001), which is attributed to the uncertainty in the MPC optimization that can also be noticed in the validation of MPC (Fig. 3). Furthermore, one caveat when using actual forecasts (as the case of NMME forecasts) is to consider the forecast skill due to the inherent uncertainty in the forecasting model, especially when dealing with extremes events (e.g., dry and wet years). A lower forecast skill can significantly reduce the value to reservoir operation, e.g., if a dry year is falsely forecasted as a wet year. Notwithstanding, the improvement of HAD operation, in the example presented, highlighted the value of streamflow forecast to adapt reservoir operation to future challenges, e.g., GERD filling or increasing demands.
The HAD (left) storage volume and (right) normalized penalty cost when using streamflow forecast (perfect, climatology, and NMME) in wet (2001) and dry (2002) years (only 1-, 3-, and 6-month lead times are shown).
Citation: Journal of Hydrometeorology 22, 5; 10.1175/JHM-D-20-0241.1
2) GERD filling and operation phases
Figure 11 shows the normalized change in penalty cost of HAD operation during GERD filling and operation as compared to the current conditions without GERD. Unsurprisingly, the penalty cost increased significantly during the GERD filling, especially at faster filling scenario. For example, when testing a 3-yr filling scenario, HAD can operate to meet the current target objectives (i.e., target storage and releases) with an average increase in the normalized optimal cost ranging between 5.1% and 3.8% for 1- and 12-month lead times, respectively (Fig. 11a). Such increase in penalty cost is attributed to less flow contribution from the Blue Nile as the GERD starts its filling (Fig. 11d) and therefore penalizes HAD operation to meet its target storage level and demands. With a slower filling scenario, e.g., 7-yr filling, streamflow forecasting can have a higher value in HAD operation as indicated by the reduction in the normalized penalty costs (Fig. 11b). A 7-yr filling scenario can improve the forecast value to HAD operation with an average reduction in the operation penalty cost of 85% (averaged for all lead times) as opposed to 3-yr filling scenario. This reduction is explained by the higher annual GERD releases with 7-yr filling approach that increases by about 60% (37 km3 compared to only 23.3 km3 with 3-yr filling scenario). During the GERD operation phase, the spatial pattern of HAD operation costs changes with higher penalty cost in the months following the rainy season upstream of GERD (September through November). The maximum operation penalty cost of HAD during GERD operation is noticed in October with an increase of 101.3 (km3)2 (normalized change in penalty cost = 0.39%) and 41.4 (km3)2 (normalized change in penalty cost = 0.16%) for 1- and 12-month lead time, respectively (Fig. 11c). Such increase follows the significant regulation in the flow downstream of GERD with higher releases in the nonflooding seasons (Fig. 11d).
(a),(b) Change in normalized penalty cost of HAD operation during GERD filling and (c) operation as compared to current conditions without GERD. The penalty cost is calculated using perfect forecast and assuming a demand trajectory of 50th percentile. (d) GERD release for different scenarios including no GERD, 3- and 7-yr GERD filling, and GERD operation (postfilling).
Citation: Journal of Hydrometeorology 22, 5; 10.1175/JHM-D-20-0241.1
Figure 12 compares the FVA as we introduce streamflow forecast at longer lead time to HAD operation under different GERD scenarios. Streamflow forecasting has higher value (i.e., higher reduction in cost) for the current conditions without GERD as compared to GERD filling scenarios. A 3-yr filling scenario of GERD has the lowest FVA with an average FVA of 2.5% (averaged for all lead times). This can be explained by the fact that a fast filling scenario will significantly drag the HAD storage to low levels and hence, introducing streamflow forecast information would minimally help in cost reduction. On the contrary, when adapting a slower filling scenario, e.g., 7 years, HAD can reap the value of streamflow forecasting, especially at shorter lead times, less than 4 months. For instance, a FVA of 6.3% can be attained when operating HAD with a 2-month lead time streamflow forecasting. During GERD operation, the streamflow forecasts tend to benefit the HAD operation at longer lead times. In addition, the forecast horizon (FVA = 0) of HAD operation increases to 9-month lead time (compared to 8 months in case of current conditions without GERD). For example, GERD operates in the future, HAD can benefit from an 8-month lead time streamflow forecasting with a cost reduction of 8.4% compared to only 1.4% for the case without GERD. Overall, maintaining the same target levels of HAD objective during GERD filling and operation would incur more operation penalty costs. This recommends revising the current HAD operating curve to cope with expected changes in upstream flow (Fig. 11d). The analysis highlighted the months with higher penalty costs (e.g., October during GERD operation), where adaptation of reservoir operation can be employed to better benefit from streamflow forecasts.
Average FVA (i.e., reduction in month-over-month optimal cost) using perfect forecast during GERD filling and operation scenarios and assuming 50th percentile demand trajectory downstream of HAD.
Citation: Journal of Hydrometeorology 22, 5; 10.1175/JHM-D-20-0241.1
5. Conclusions
This study investigated the value of long-term streamflow forecast to improve reservoir operation through a Forecast-Based Adaptive Reservoir Operation (FARO) framework. Forecasting streamflow at subseasonal (weeks) to seasonal time scales (months ahead) has a prominent role in management of water resources including reservoir operation and management, flood risk planning, and water allocation for crop irrigation. While the value of inflow forecasts in reservoir operation has been capitalized in previous studies, understanding of their precise contribution to adaptive reservoir operation in transboundary basins is limited. Our FARO approach introduces a formalized procedure to quantify the forecast value of long-term streamflow forecasts in adaptive reservoir operation. A forecast-based approach outperforms a traditional static operating policy, with the added benefit from the forecast information. The FARO approach was applied to the High Aswan Dam (HAD) in the Nile River basin to assess the value of streamflow forecasts at different lead time and under expected challenges from upstream planned dams [e.g., Grand Ethiopian Renaissance Dam (GERD)].
Our key conclusions are as follows:
The FARO modeling framework (i.e., MPC and VIC models) shows a high accuracy in simulating the actual HAD reservoir and capturing the interannual and intra-annual storage and flow variations. In our FARO framework, a lead-time-dependent bias correction method resulted in highly skillful NMME-based streamflow forecasts at Khartoum station with an average NSE and correlation coefficient of 0.84 and 0.92, respectively. The ability of both MPC and VIC models to simulate existing reservoir operation and upstream inflow forecasts is critical to evaluating the viability of the forecast information in adapting the HAD operation.
The forecast value added (FVA) to HAD operation is more pronounced in the months following the flooding season (October–December). Issuing forecasting in these months is key to HAD operation and can significantly improve the decisions to release or store water in the next water year. When using actual forecasts from NMME models from October through December, HAD operation was improved with an average gain in forecast skill ranging from 16% (1-month lead time) to 231% (12-month lead time) as compared to the climatology-based forecasts. The NMME forecast skill does not decrease uniformly with lead time, and the selection of a skillful lead time varies as a function of forecast issuing month (e.g., for winter months, higher skill is attained at shorter lead times).
The FARO framework has shown its value in adapting reservoir operation when implemented under different conditions, including varying downstream demands, streamflow climatology during dry versus wet years, and upstream structural features introduced into the water system. The filling of GERD, especially at faster approach, can significantly drag the HAD storage to low levels. During its operation phase, GERD will regulate the HAD inflow throughout the year, i.e., less inflow in the summer months. Hence, introducing streamflow forecast would minimally help to maintain current target objectives of HAD operation and therefore result in higher operation costs. To improve the forecast value while considering GERD operation, deriving a new adaptive operating policy for HAD is recommended.
Our FARO example on HAD and GERD operation reveals the potential of FARO framework to evaluate the value of streamflow forecast in operation of other dams in the basin, e.g., Sennar, Roseires, and Merowe Dams in Sudan. Furthermore, given the advantage of being routinely produced and publicly accessible, NMME forecasts can be implemented operationally to guide water release decisions. Integrating streamflow forecasts into reservoir operations can support water managers and various stakeholders in the Nile region. For a country like Egypt, a FARO framework can be operationalized into a decision support system to monitor reservoir operations, i.e., water storage and release patterns, by upstream riparian countries. Operationalizing long-term streamflow forecasting will have the potential to help negotiate the filling and operation decisions of new dams by estimating annual expectations of streamflow. A major challenge when implementing forecasts in real-time operational applications is the uncertainty involved in streamflow forecasts (Zhao et al. 2011; Chen et al. 2016). Considering forecast uncertainty through model ensembles can compensate for the loss in performance due to forecast inaccuracy. Our study demonstrated that streamflow forecasts based on NMME models can potentially improve real-time reservoir operations. To further enhance the NMME forecast skill, future work can also benefit from hybrid models that integrate both dynamical climate forecasts and statistical or data-driven methods (Slater and Villarini 2018). For example, a hybrid approach can learn from GCM physical-based forecasts to inform the predictors in statistical forecasting methods (see for example, Robertson et al. 2013 and Humphrey et al. 2016).
Increases in the frequency and intensity of extreme weather events are anticipated to alter streamflow patterns and affect reservoir operation (Asadieh and Krakauer 2015; Ehsani et al. 2017). An extreme event that occurs downstream of an existing dam can lead to a severe flooding or drought if no prior adaptation is implemented to the dam releases. This would imply, for example, releasing less water from the reservoir before it rains. Using short-term forecasts can help to predict downstream extreme events and allow employing an adaptive reservoir operation. In October 2015, Egypt has been subjected to extreme precipitation events in the Nile delta (downstream of HAD) that has resulted in severe flooding of some regions (Zevenbergen et al. 2017). One reason for downstream flooding is the failure of HAD to adapt its operation in a coordinated way that account for downstream precipitation events. The FARO framework can be further extended to benefit from the application of existing numerical forecast models, e.g., NCEP Global Forecast System (GFS), to provide weather forecasts at weekly to subdaily scales. This would pave the way to derive adaptive operating policy of existing dams while exploiting the value of both long- and short-term forecasts.
Acknowledgments
This work supported by NASA Surface Water Ocean Topography mission Science Team Grant (NNX16AQ54G), NSF EAR 1740042 titled INFEWS/T1: Linking Current and Future Hydrologic Change to Hydropower, Human Nutrition, and Livelihoods in the Lower Mekong Basin. Additional support to second author from NASA Applied Sciences SERVIR Grant 80NSSC20K0152 is also gratefully acknowledged.
Data availability statement
All data and models used in this study are publicly available. The VIC hydrologic model is available from https://vic.readthedocs.io/en/master/. CHIRPS satellite precipitation data are available from
REFERENCES
Ahmad, S. K., and F. Hossain, 2020: Maximizing energy production from hydropower dams using short-term weather forecasts. Renewable Energy, 146, 1560–1577, https://doi.org/10.1016/j.renene.2019.07.126.
Ahmadi, M., O. B. Haddad, and H. A. Loáiciga, 2015: Adaptive reservoir operation rules under climatic change. Water Resour. Manage., 29, 1247–1266, https://doi.org/10.1007/s11269-014-0871-0.
Anghileri, D., N. Voisin, A. Castelletti, F. Pianosi, and B. Nijssen, and D. P. Lettenmaier, 2016: Value of long-term streamflow forecasts to reservoir operations for water supply in snow-dominated river catchments. Water Resour. Res., 52, 4209–4225, https://doi.org/10.1002/2015WR017864.
Asadieh, B., and N. Y. Krakauer, 2015: Global trends in extreme precipitation: Climate models versus observations. Hydrol. Earth Syst. Sci., 19, 877–891, https://doi.org/10.5194/hess-19-877-2015.
Bakken, T. H., I. S. Modahl, H. L. Raadal, A. A. Bustos, and S. Arnøy, 2016: Allocation of water consumption in multipurpose reservoirs. Water Policy, 18, 932–947, https://doi.org/10.2166/wp.2016.009.
Biemans, H., I. Haddeland, P. Kabat, F. Ludwig, R. W. A. Hutjes, J. Heinke, W. von Bloh, and D. Gerten, 2011: Impact of reservoirs on river discharge and irrigation water supply during the 20th century. Water Resour. Res., 47, W03509, https://doi.org/10.1029/2009WR008929.
Block, P., and L. Goddard, 2012: Statistical and dynamical climate predictions to guide water resources in Ethiopia. J. Water Resour. Plann. Manage., 138, 287–298, https://doi.org/10.1061/(ASCE)WR.1943-5452.0000181.
Block, P., and B. Rajagopalan, 2007: Interannual variability and ensemble forecast of Upper Blue Nile Basin Kiremt season precipitation. J. Hydrometeor., 8, 327–343, https://doi.org/10.1175/JHM580.1.
Breckpot, M., O. M. Agudelo, and B. De Moor, 2013: Flood control with model predictive control for river systems with water reservoirs. J. Irrig. Drain. Eng., 139, 532–541, https://doi.org/10.1061/(ASCE)IR.1943-4774.0000577.
Brown, C. M., and Coauthors, 2015: The future of water resources systems analysis: Toward a scientific framework for sustainable water management. Water Resour. Res., 51, 6110–6124, https://doi.org/10.1002/2015WR017114.
Candogan Yossef, N., R. Van Beek, A. Weerts, H. Winsemius, and M. F. Bierkens, 2017: Skill of a global forecasting system in seasonal ensemble streamflow prediction. Hydrol. Earth Syst. Sci., 21, 4103–4114, https://doi.org/10.5194/hess-21-4103-2017.
Chen, L., V. P. Singh, W. Lu, J. Zhang, J. Zhou, and S. Guo, 2016: Streamflow forecast uncertainty evolution and its effect on real-time reservoir operation. J. Hydrol., 540, 712–726, https://doi.org/10.1016/j.jhydrol.2016.06.015.
Chipman, J. W., 2019: A multisensor approach to satellite monitoring of trends in lake area, water level, and volume. Remote Sens., 11, 158, https://doi.org/10.3390/rs11020158.
Delaney, C. J., and Coauthors, 2020: Forecast informed reservoir operations using ensemble streamflow predictions for a multipurpose reservoir in Northern California. Water Resour. Res., 56, e2019WR026604, https://doi.org/10.1029/2019WR026604.
Ehsani, N., C. J. Vörösmarty, B. M. Fekete, and E. Z. Stakhiv, 2017: Reservoir operations under climate change: Storage capacity options to mitigate risk. J. Hydrol., 555, 435–446, https://doi.org/10.1016/j.jhydrol.2017.09.008.
Eldardiry, H., and F. Hossain, 2019: Understanding reservoir operating rules in the transboundary Nile River basin using macroscale hydrologic modeling with satellite measurements. J. Hydrometeor., 20, 2253–2269, https://doi.org/10.1175/JHM-D-19-0058.1.
Eldardiry, H., and F. Hossain, 2020: A blueprint for adapting high Aswan Dam operation in Egypt to challenges of filling and operation of the Grand Ethiopian Renaissance Dam. J. Hydrol., 125708, https://doi.org/10.1016/j.jhydrol.2020.125708.
Eldardiry, H., and F. Hossain, 2021: Evaluating the hydropower potential of the Grand Ethiopian Renaissance Dam. J. Renewable Sustainable Energy, 13, 024501, https://doi.org/10.1063/5.0028037.
Feng, M., P. Liu, S. Guo, L. Shi, C. Deng, and B. Ming, 2017: Deriving adaptive operating rules of hydropower reservoirs using time-varying parameters generated by the EnKF. Water Resour. Res., 53, 6885–6907, https://doi.org/10.1002/2016WR020180.
Galelli, S., A. Goedbloed, D. Schwanenberg, and P. J. van Overloop, 2014: Optimal real-time operation of multipurpose urban reservoirs: Case study in Singapore. J. Water Resour. Plann. Manage., 140, 511–523, https://doi.org/10.1061/(ASCE)WR.1943-5452.0000342.
Georgakakos, K. P., and N. E. Graham, 2008: Potential benefits of seasonal inflow prediction uncertainty for reservoir release decisions. J. Appl. Meteor. Climatol., 47, 1297–1321, https://doi.org/10.1175/2007JAMC1671.1.
Giuliani, M., M. Zaniolo, A. Castelletti, G. Davoli, and P. Block, 2019: Detecting the state of the climate system via artificial intelligence to improve seasonal forecasts and inform reservoir operations. Water Resour. Res., 55, 9133–9147, https://doi.org/10.1029/2019WR025035.
Golembesky, K., A. Sankarasubramanian, and N. Devineni, 2009: Improved drought management of Falls Lake Reservoir: Role of multimodel streamflow forecasts in setting up restrictions. J. Water Resour. Plann. Manage., 135, 188–197, https://doi.org/10.1061/(ASCE)0733-9496(2009)135:3(188).
Hamlet, A. F., and D. P. Lettenmaier, 1999: Columbia River streamflow forecasting based on ENSO and PDO climate signals. J. Water Resour. Plann. Manage., 125, 333–341, https://doi.org/10.1061/(ASCE)0733-9496(1999)125:6(333).
Hamlet, A. F., D. Huppert, and D. P. Lettenmaier, 2002: Economic value of long-lead streamflow forecasts for Columbia River hydropower. J. Water Resour. Plann. Manage., 128, 91–101, https://doi.org/10.1061/(ASCE)0733-9496(2002)128:2(91).
Humphrey, G. B., M. S. Gibbs, G. C. Dandy, and H. R. Maier, 2016: A hybrid approach to monthly streamflow forecasting: Integrating hydrological model outputs into a Bayesian artificial neural network. J. Hydrol., 540, 623–640, https://doi.org/10.1016/j.jhydrol.2016.06.026.
Jabbari, A., and D. H. Bae, 2020: Improving ensemble forecasting using total least squares and lead-time dependent bias correction. Atmosphere, 11, 300, https://doi.org/10.3390/atmos11030300.
Jasperse, J., and Coauthors, 2017: Preliminary viability assessment of Lake Mendocino forecast informed reservoir operations. Center for Western Weather and Water Extremes Tech. Rep., 75 pp., https://cw3e.ucsd.edu/FIRO_docs/FIRO_PVA.pdf
Kirtman, B. P., and Coauthors, 2014: The North American multimodel ensemble: Phase-1 seasonal-to-interannual prediction; phase-2 toward developing intraseasonal prediction. Bull. Amer. Meteor. Soc., 95, 585–601, https://doi.org/10.1175/BAMS-D-12-00050.1.
Labadie, J. W., 2004: Optimal operation of multireservoir systems: State-of-the-art review. J. Water Resour. Plann. Manage., 130, 93–111, https://doi.org/10.1061/(ASCE)0733-9496(2004)130:2(93).
Lettenmaier, D. P., 2017: Observational breakthroughs lead the way to improved hydrological predictions. Water Resour. Res., 53, 2591–2597, https://doi.org/10.1002/2017WR020896.
Liang, X., D. P. Lettenmaier, E. F. Wood, and S. J. Burges, 1994: A simple hydrologically based model of land surface water and energy fluxes for general circulation models. J. Geophys. Res., 99, 14 415–14 428, https://doi.org/10.1029/94JD00483.
Liang, X., E. F. Wood, and D. P. Lettenmaier, 1996: Surface soil moisture parameterization of the VIC-2L model: Evaluation and modification. Global Planet. Change, 13, 195–206, https://doi.org/10.1016/0921-8181(95)00046-1.
Ma, F., X. Yuan, and A. Ye, 2015: Seasonal drought predictability and forecast skill over China. J. Geophys. Res. Atmos., 120, 8264–8275, https://doi.org/10.1002/2015JD023185.
Maurer, E. P., and D. P. Lettenmaier, 2004: Potential effects of long-lead hydrologic predictability on Missouri River main-stem reservoirs. J. Climate, 17, 174–186, https://doi.org/10.1175/1520-0442(2004)017<0174:PEOLHP>2.0.CO;2.
Misra, V., and H. Li, 2014: The seasonal climate predictability of the Atlantic Warm Pool and its teleconnections. Geophys. Res. Lett., 41, 661–666, https://doi.org/10.1002/2013GL058740.
Moussa, A. M. A., 2018: Dynamic operation rules of multi-purpose reservoir for better flood management. Alex. Eng. J., 57, 1665–1679, https://doi.org/10.1016/j.aej.2017.03.012.
Mulat, A. G., and S. A. Moges, 2014: Assessment of the impact of the Grand Ethiopian Renaissance Dam on the performance of the High Aswan Dam. J. Water Resour. Prot., 6, 583–598, https://doi.org/10.4236/jwarp.2014.66057.
Munia, H., J. H. A. Guillaume, N. Mirumachi, M. Porkka, Y. Wada, and M. Kummu, 2016: Water stress in global transboundary river basins: Significance of upstream water use on downstream stress. Environ. Res. Lett., 11, 014002, https://doi.org/10.1088/1748-9326/11/1/014002.
Nayak, M. A., J. D. Herman, and S. Steinschneider, 2018: Balancing flood risk and water supply in California: Policy search integrating short-term forecast ensembles with conjunctive use. Water Resour. Res., 54, 7557–7576, https://doi.org/10.1029/2018WR023177.
Robertson, D. E., P. Pokhrel, and Q. J. Wang, 2013: Improving statistical forecasts of seasonal streamflows using hydrological model output. Hydrol. Earth Syst. Sci., 17, 579–593, https://doi.org/10.5194/hess-17-579-2013.
Rodell, M., J. S. Famiglietti, D. N. Wiese, J. T. Reager, H. K. Beaudoing, F. W. Landerer, and M. H. Lo, 2018: Emerging trends in global freshwater availability. Nature, 557, 651–659, https://doi.org/10.1038/s41586-018-0123-1.
Sadek, N., and M. Aziz, 2005: Flood management of Lake Nasser after the new Toshka Barrages construction. Ninth Int. Water Technology Conf., Sharm El-Sheikh, Egypt, Water Technology Association, 1009–1025.
Shumilova, O., K. Tockner, M. Thieme, A. Koska, and C. Zarfl, 2018: Global water transfer megaprojects: A potential solution for the water-food-energy nexus? Front. Environ. Sci., 6, 150, https://doi.org/10.3389/fenvs.2018.00150.
Slater, L. J., and G. Villarini, 2018: Enhancing the predictability of seasonal streamflow with a statistical-dynamical approach. Geophys. Res. Lett., 45, 6504–6513, https://doi.org/10.1029/2018GL077945.
Slater, L. J., G. Villarini, and A. A. Bradley, 2019: Evaluation of the skill of North-American multi-model ensemble (NMME) global climate models in predicting average and extreme precipitation and temperature over the continental USA. Climate Dyn., 53, 7381–7396, https://doi.org/10.1007/s00382-016-3286-1.
Taye, M. T., and P. Willems, 2012: Temporal variability of hydroclimatic extremes in the Blue Nile basin. Water Resour. Res., 48, W03513, https://doi.org/10.1029/2011WR011466.
Tesemma, Z. K., Y. A. Mohamed, and T. S. Steenhuis, 2010: Trends in rainfall and runoff in the Blue Nile Basin: 1964–2003. Hydrol. Processes, 24, 3747–3758, https://doi.org/10.1002/hyp.7893.
Turner, S. W., J. C. Bennett, D. E. Robertson, and S. Galelli, 2017: Complex relationship between seasonal streamflow forecast skill and value in reservoir operations. Hydrol. Earth Syst. Sci., 21, 4841–4859, https://doi.org/10.5194/hess-21-4841-2017.
Wan, W., J. Zhao, J. R. Lund, T. Zhao, X. Lei, and H. Wang, 2016: Optimal hedging rule for reservoir refill. J. Water Resour. Plann. Manage., 142, 04016051, https://doi.org/10.1061/(ASCE)WR.1943-5452.0000692.
Wheeler, K. G., J. W. Hall, G. M. Abdo, S. J. Dadson, J. R. Kasprzyk, R. Smith, and E. A. Zagona, 2018: Exploring cooperative transboundary river management strategies for the Eastern Nile Basin. Water Resour. Res., 54, 9224–9254, https://doi.org/10.1029/2017WR022149.
Woldemeskel, F., D. McInerney, J. Lerat, M. Thyer, D. Kavetski, D. Shin, N. Tuteja, and G. Kuczera, 2018: Evaluating post-processing approaches for monthly and seasonal streamflow forecasts. Hydrol. Earth Syst. Sci., 22, 6257–6278, https://doi.org/10.5194/hess-22-6257-2018.
Wood, A. W., and J. C. Schaake, 2008: Correcting errors in streamflow forecast ensemble mean and spread. J. Hydrometeor., 9, 132–148, https://doi.org/10.1175/2007JHM862.1.
Yang, G., S. Guo, P. Liu, and P. Block, 2020: Integration and evaluation of forecast-informed multiobjective reservoir operations. J. Water Resour. Plann. Manage., 146, 04020038, https://doi.org/10.1061/(ASCE)WR.1943-5452.0001229.
Ye, A., Q. Duan, X. Yuan, E. F. Wood, and J. Schaake, 2014: Hydrologic post-processing of MOPEX streamflow simulations. J. Hydrol., 508, 147–156, https://doi.org/10.1016/j.jhydrol.2013.10.055.
Yuan, X., E. F. Wood, and Z. Ma, 2015: A review on climate-model-based seasonal hydrologic forecasting: Physical understanding and system development. Wiley Interdiscip. Rev.: Water, 2, 523–536, https://doi.org/10.1002/WAT2.1088.
Zarfl, C., A. E. Lumsdon, J. Berlekamp, L. Tydecks, and K. Tockner, 2015: A global boom in hydropower dam construction. Aquat. Sci., 77, 161–170, https://doi.org/10.1007/s00027-014-0377-0.
Zaroug, M. A. H., E. A. B. Eltahir, and F. Giorgi, 2013: Droughts and floods over the upper catchment of the Blue Nile and their connections to the timing of El Niño and La Niña events. Hydrol. Earth Syst. Sci., 18, 1239–1249, https://doi.org/10.5194/HESS-18-1239-2014.
Zevenbergen, C., B. Bhattacharya, R. A. Wahaab, W. A. I. Elbarki, T. Busker, and C. S. Rodriguez, 2017: In the aftermath of the October 2015 Alexandria flood challenges of an Arab city to deal with extreme rainfall storms. Nat. Hazards, 86, 901–917, https://doi.org/10.1007/s11069-016-2724-z.
Zhao, T., X. Cai, and D. Yang, 2011: Effect of streamflow forecast uncertainty on real-time reservoir operation. Adv. Water Resour., 34, 495–504, https://doi.org/10.1016/j.advwatres.2011.01.004.
Zhao, T., D. Yang, X. Cai, J. Zhao, and H. Wang, 2012: Identifying effective forecast horizon for real-time reservoir operation under a limited inflow forecast. Water Resour. Res., 48, W01540, https://doi.org/10.1029/2011WR010623.