23

Q IWA Publishing 2008 Journal of Hydroinformatics | 10.1 | 2008

Multistage stochastic linear programming model for daily coordinated multi-reservoir operation
Yongdae Lee, Sheung-Kown Kim and Ick Hwan Ko

ABSTRACT
Operation planning for a coordinated multi-reservoir is a complex and challenging task due to the inherent uncertainty in inow. In this study, we suggest the use of a new, multi-stage and scenariobased stochastic linear program with a recourse model incorporating the meteorological weather prediction information for daily, coordinated, multi-reservoir operation planning. Stages are dened as prediction lead-time spans of the weather prediction system. The multi-stage scenarios of the stochastic model are formed considering the reliability of rainfall prediction for each lead-time span. Future inow scenarios are generated by a rainfallrunoff model based on the rainfall forecast. For short-term stage (2 days) scenarios, the regional data assimilation and prediction system (RDAPS) information is employed, and for mid-term stage (more than 2 days) scenarios, precipitation from the global data assimilation and prediction system (GDAPS) is used as an input for the rainfallrunoff model. After the 10th day (third stage), the daily historical rainfall data are used following the ensemble streamow prediction (ESP) procedure. The model is applied to simulate the daily reservoir operation of the Nakdong River basin in Korea in a real-time operational environment. The expected benet of the stochastic model is markedly superior to that of the deterministic model with average rainfall information. Our study results conrm the effectiveness of the stochastic model in real-time operation with meteorological forecasts and the presence of inow uncertainty.
Key words
Sheung-Kown Kim (corresponding author) Division of Information Management Engineering, Korea University, 5-1 Anam-Dong, Sungbook-Gu, Seoul 136-701, Korea Tel.: +82 2 3290 3385 Fax: +82 2929 5888 E-mail: [email protected] Ick Hwan Ko Korea Institute of Water and Environment, Korea Water Resources Corporation, 462-1 Jeonmin-dong, Yuseong-Gu, Daejeon 305-730, Korea Yongdae Lee Graduate School of Information Management and Security, Korea University, 5-1 Anam-Dong, Sungbook-Gu, Seoul 136-701, Korea

| daily reservoir operation, ESP, GDAPS, RDAPS, scenario generation, stochastic

programming

INTRODUCTION
Coordinated, multi-reservoir system, operation planning is essentially the task of the optimum allocation of water among the dams in the system to determine the storage level and release quantity of each reservoir over the planning period. Most of the analytic frameworks for multi-reservoir system operation involving deterministic optimization and simulation models were established in the 1970s and early 1980s (Yazicigil et al. 1983; Can & Houck 1984). Yeh (1985), Wurbs (1993) and Labadie (2004) provided extensive lists of references on the use of simulation and optimization methods in reservoir-system operation. However, we cannot ignore uncertainties such as the uncertainty of parameter values, input data, uncertainty associated with projections
doi: 10.2166/hydro.2008.007

of future demands, or uncertainty of inow. Sensitivity analysis or parametric programming techniques were used to analyze the risk and uncertainty of parameter or input data. These techniques can help estimate the extent to which we need to reduce these uncertainties. In this study, we use a scenario-based, stochastic linear programming approach to cope with the inow uncertainty because it is a clear source of uncertainty. We assume that demand uncertainties in water supply planning are less signicant than the inow uncertainty. Setting water demand as a constraint is sufcient, since we know that a hydrologic basin cannot deliver more water than is available, and demand requirements are usually determined by the historical records, and set by contracts.

24

Y. Lee et al. | Multistage stochastic linear programming model

Journal of Hydroinformatics | 10.1 | 2008

Many stochastic optimization models such as stochastic dynamic programming (SDP) (Loucks et al. 1981; Yakowitz 1982; Stedinger et al. 1984), chance-constrained programming (CCP) (ReVelle et al. 1969; Houck 1979), two-stage stochastic linear programming (SLP) with recourse (Sei & Hipel 2001) and multi-stage stochastic programming (Pereira & Pinto 1985; Jacobs et al. 1995; Watkins et al. 2000; Kracman et al. 2006) have been studied to deal with inow uncertainty. SDP models can generate effective and feasible release policies for every possible reservoir storage state in each month. However, SDP usually requires discrete state variables and is limited by the so-called curse of dimensionality (Yeh 1985; Pereira & Pinto 1985). CCP is focussed on the reliability of the solution, which sets a minimum requirement on the probability of constraint satisfaction. This problem can be converted to a general deterministic linear programming problem, but it is often non-convex unless the right-hand side has a log-concave multivariate probability density function (Prekopa 1995). Hence, SLP seems to be the most practical approach to deal with the inow uncertainty. Two-stage SLP with recourse has been applied for longterm planning under uncertainty since it was rst proposed by Dantzig (1955) and Beale (1955). In the rst stage, the non-anticipative planning level decisions (here and now) are usually determined independently of the subsequent variations in random scenarios, and then the control or operational decisions are made with consideration for the feasibility in the second stage, which is called a recourse action. The two-stage SLP can be expanded to a multi-stage problem. A nite number of the scenarios, corresponding to sequences of realizations of random variables at each stage, must be specied to apply the multi-stage SLP with recourse (Dupacova et al. 2000). The multi-stage SLP for reservoir operation seems to be a viable approach if we can identify stages at which a decision is required, and when release decisions taken at any stage of the decision process do not depend on future realizations of the inows or on future release decisions. However, it has the shortcoming that the problem size of the deterministic equivalent of multi-stage SLP is usually very large and grows exponentially with the number of decision stages and linearly with the number of scenarios at each stage (Sei & Hipel 2001). To overcome these difculties, many solution methods such as the L-shaped algorithm (Van Slyke & Wets 1969; Wets 1988), Benders decomposition approaches (Jacobs et al. 1995),

Dantzig Wolfe decomposition (Dantzig & Madansky 1961) and interior-point methods (Ponnambalam et al. 1989; Sei & Hipel 2001) have been proposed. However, with the increasing computing power and development of many solution packages (CPLEX, MSLiP-OSL, OSL-SP, etc.), the choice or construction of a suitable model, that takes into account the nature of the real-life problem, input data characteristics, software availability and computer technology, has become more important than algorithmic development (Dupacova et al. 2000). Most of the SLP models for operation planning of coordinated multi-reservoirs are applied to long-term operational planning (Pereira & Pinto 1985; Jacobs et al. 1995; Watkins et al. 2000; Kracman et al. 2006). Sei & Hipel (2001) suggested the use of two-stage SLP with a simple recourse model, using the interior-point method for the monthly operation of the Great Lakes. In this study, we suggest the use of multi-stage SLP with recourse for daily reservoir operation. Although the suggested model looks similar to other SLP models since it follows the general structure of multi-stage SLP, the problem nature of our model is quite different from others. Our model is unique in that the stages are dened as prediction lead-time spans of the weather prediction system, multiple objectives are utilized, and the model is designed for direct application to real-time operation. Most of the studies on daily operation or real-time operation use deterministic approaches (Yazicigil et al. 1983; Mujumdar & Ramesh 1997; Eschenbach et al. 2001; Turgeon 2005; Kim et al. 2005). However, deterministic models are incapable of considering hedging decisions and the risk of water supply shortage or ooding against the inherent uncertainty of the inow. In this study, we attempted to use SLP for the daily real-time operation simulation of multiple reservoirs in the Nakdong River Basin in Korea, and investigated possibility of using SLP for real-time daily operation. The probable benets are evaluated and the results are summarized in the fourth section. Jacobs et al. (1995) suggested the use of a stochastic optimization model for medium-term optimal operation planning for PG&Es hydroelectric generation. Unlike ours, there is no report on the real benets that might be obtained as a result of implementing the multi-stage SLP for medium-term planning operation. They formulated a scenario tree and forecasted the streamow by using the characteristics of snowpack; furthermore, they

25

Y. Lee et al. | Multistage stochastic linear programming model

Journal of Hydroinformatics | 10.1 | 2008

indeed dened the stages based on the accuracy of streamow forecasting. However, their model is designed to nd optimal scheduling of hydroelectric generation and is unsuitable for multi-reservoir systems such as the Korean hydrologic basin, where water supply is the primary purpose and snowpack is not a signicant factor. Since we wanted to have a multi-stage SLP model for real-time daily reservoir operation planning, we built the multi-stage scenario tree considering both the accuracy of short-term, mid-term or long-term meteorological forecasts, but unlike Watkins et al. (2000) or Kracman et al. (2006), we generated streamow scenarios using the rainfall runoff model with meteorological forecasting information in order to preserve the spatial and serial correlations. In this study, we formulate a multiperiod, multi-stage, SLP model to derive an efcient daily release plan for each reservoir and we suggest the use of it for daily, coordinated, multi-reservoir operation planning. The multi-stage scenarios for the stochastic model are formulated considering the reliability of the future inow prediction for a specied forecasting leadtime. The future inows are generated by a rainfall runoff model based on the rainfall forecasts. In Korea, we rely on two climate modelling systems which are both operated by the Korea Meteorological Administration (KMA). One is the global data assimilation and prediction system (GDAPS) and the other is the regional data assimilation and prediction system (RDAPS). GDAPS is a numerical climate modelling system designed to predict weather condition by dividing the global atmospheric domain into vertical levels and horizontal resolutions using ocean wave prole and synoptic observations, including satellite retrieval data. RDAPS is also a numerical modelling system with high resolution 3D grids designed to predict regional weather condition, using GDAPS as its temporal boundary condition. It uses more enriched and detailed regional asynoptic data, such as wind and moisture prole, temperature, automatic weather station and radar data, to predict local precipitation. The type of rainfall prediction model for each prediction lead-time is selected based upon the viability and accuracy of the prediction. For short-term (2 days) rainfall forecasting, RDAPS is applied. The GDAPS is used for mid-term (10 days) rainfall forecasts. After the tenth day (third stage), the daily historical rainfall data are applied. The model becomes essentially a large-scale, linear programming model

(comprising over 180,000 columns and 120,000 rows), and was solved using the mathematical programming solver, CPLEX 9.0. In the following section, the model formulation is presented. The third section presents a case study in which a simulation of daily reservoir operation for the Nakdong River Basin in Korea under uncertainty is performed under the realtime operational environment. In the fourth section, we evaluate the value of the stochastic model based on Birges measure (Birge & Louveaux 1997). The conclusions and future extensions are presented in the nal section.

MODEL FORMULATION
To derive an efcient daily release plan for each reservoir, a multi-period, multi-stage, SLP with recourse model is formulated. The planning decision involves determining the daily release plan, while the scenario-dependent control decisions on storage, hydroelectric energy generation, water supply, etc., constitute recourse actions to minimize any infeasibility and maximize the operational benet. The multi-stage scenarios for the stochastic model are generated by a rainfall runoff model based on the meteorological rainfall forecasts with consideration for the accuracy of future rainfall prediction for each forecasting lead-time span. The objective function of the model is dened as the weighted sum of multiple objectives. The weighting factors of each objective are assigned preemptively. The constraints sets include ow conservation of the multi-period, dynamic network ow, physical limitation of nodes and arcs, and non-anticipativity constraints.

Multi-stage inow scenario tree In order to elaborate the inow uncertainty in the decision process explicitly, scenario-based multi-stage stochastic optimization is applied. A multi-stage scenario tree enables us to keep track of uncertain decision states following the model of rainfall prediction. The accuracy of two-day rainfall forecasts has been improved by the increased meteorological information available due to the signicantly better understanding of the climatic system. Accordingly, the accuracy of the

26

Y. Lee et al. | Multistage stochastic linear programming model

Journal of Hydroinformatics | 10.1 | 2008

streamow forecast has also been improved with the advance of computer modelling technology in hydrological science. However, rainfall forecasting is still not reliable if the prediction lead-time is longer than 10 days. The multi-stage inow scenario tree is built considering both the accuracy and reliability of short-term meteorological forecasts. The accuracy of the streamow forecasting is dependent upon the quality of the rainfall forecasting for different lead-time spans. In the model, stages are dened as time spans corresponding to the available meteorological forecast information, as shown in Figure 1. In the rst stage (rst 2 days), the rainfall forecast using information from RDAPS is applied. As RDAPS provides only 2 days forecasts twice a day, we set 2 days as the time span of the rst stage. In the second stage (from the third to the tenth day), the rainfall forecasts from the GDAPS are linked to the previous data, and the historical minimum and maximum rainfall information are supplemented because the accuracy of GDAPS forecasts are not sufcient. As GDAPS forecasts can provide meteorological information for 10 days once a day, 10 days were set as the end of the time span of the second stage. After the tenth day (third stage), the rainfall forecasts are unreliable so the observed daily precipitation data from every year available in the historical records were concatenated to obtain the ESP (Day 1985).

Although the multiple scenarios based on the historical inow after the tenth day in the third stage are not reliable, it is crucial for non-anticipating decisions to attain the nal storage target, taking into account the risk of serious ooding or drought in the sequential decisions. For the case of the realtime operation, only the rst-stage decision is adopted with the good prospect of achieving the nal storage target for each streamow scenario. Therefore, the models planning horizon is designed from today to the end of the month, and more than 10 days of multiple scenarios are required. The statistics of the 2-day RDAPS rainfall forecasts are compared with automatic weather system (AWS) observations in Table 1 (KWRA (Korea Water Resources Association) 2003). In Table 1, the average correlation coefcient (CC) between RDAPS and AWS is 0.7358, which conrms the appropriateness of RDAPS forecasts for water resources planning. The statistics of the 10-day GDAPS rainfall forecasts

Table 1

Statistics of RDAPS and AWS

Forecasting time (UTC)

AWS (mm)

RDAPS (mm)

CC

RMSE

2003.07.18.00 2003.07.18.12 2003.07.19.00 2003.07.19.12 2003.07.20.00 2003.07.20.12 2003.07.21.00 2003.07.21.12 2003.07.22.12 2003.07.23.00 2003.07.23.12 2003.07.24.00 2003.07.24.12 Average

1.9007 0.4702 0.2666 0.3493 2.5275 4.7120 7.5025 7.1726 1.7903 0.0590 0.0439 0.0439 0.0479 2.0682

1.9855 0.5153 0.7106 0.6524 0.8098 4.7120 6.0338 6.2250 0.4587 0.0635 0.0505 0.3362 0.1087 1.7432

0.9250 0.8226 0.2295 0.8784 0.6271 0.7707 0.2743 0.7954 0.9925 0.6054 0.8466 0.9251 0.8723 0.7358

2.0240 0.4691 1.0305 0.5558 5.6127 4.8136 8.8929 4.6269 3.6038 0.1286 0.0619 0.7330 0.1822 2.5181

Figure 1

The multi-stage scenario tree.

27

Y. Lee et al. | Multistage stochastic linear programming model

Journal of Hydroinformatics | 10.1 | 2008

Table 2

Statistics of GDAPS and AWS

Water Resources Corporation (KOWACO). The rainfall forecasts have a spatial correlation. The streamow generated based on the RRFS rainfall runoff simulation, using the observed rainfall data for one month during the warm-up period, is found to preserve the serial and spatial correlations. For real-time operation, the necessary component is the rst-stage decision which is determined independently of future hydrologic events but the feasibility of the decision must retain consideration for the variation in future hydrologic scenarios. Nonanticipativity constraints in the model necessitate that decisions at the divergence points in the scenario tree must coincide for the scenarios which have a common path up

Forecasting time (UTC) AWS (mm) GDAPS (mm) CC RMSE

2003.07.13.12 2003.07.18.12 2003.07.21.12 2003.08.29.12 Average

5.2513 7.2818 5.7626 4.3854 5.6687

3.3185 9.5823 9.4032 0.2678 5.6429

0.6321 0.7542 0.6325 0.5478 0.6416

7.3821 6.3851 7.8541 7.8845 7.3764

are summarized in Table 2 (KWRA 2003). In Table 2, the average CC between GDAPS and AWS is 0.6416, which means the GDAPS forecasts are less reliable than the RDAPS forecasts, but is still useful information for mid-term forecasts. The procedures to generate the inow scenario using the rainfall forecasts information are represented in Figure 2. The rainfall forecast information was inputted into the rainfall runoff model to generate a single streamow forecast time series. The rainfall runoff model is the rainfall runoff forecasting system (RRFS) which was developed based on the streamow synthesis and reservoir regulation model (SSARR) (US Army Corps of Engineers 2006) by the Water Resources Research Institute of Korea

to that point (Kracman et al. 2006). This ensures that the release decisions are made independently of the unknown hydrologic random event. Meanwhile, the future control decisions for storage and hydropower generation in the recourse action are specic to each inow scenario in order to ensure the feasibility of the decision. In real-time operation, the rst-stage decisions on the release of each reservoir are applied. The next day, the rainfall forecasts are updated and the process of the prediction by multi-stage and scenario-based, stochastic linear program with a recourse model and the succeeding simulation run is repeated until the end of the month, as was suggested by Kim et al. (2000).

Figure 2

The forecasted streamow scenario generated by RRFS model based on the rainfall forecast.

28

Y. Lee et al. | Multistage stochastic linear programming model

Journal of Hydroinformatics | 10.1 | 2008

Framework of the model The deterministic equivalent of our multi-stage SLP conceptual model is written as follows: min c1 T x1
N2 X k2 2

vectors (the release of each reservoir) of each path kL at the Lth stage in the scenario tree. The detailed model is presented in the appendix. This multi-stage SLP model can be so large that special
N3 X k3 N 2 1

pk2 ck2T xk2

pk3 ck3T xk3

solution algorithms like the L-shaped method may be 1 2 required (Van Slyke & Wets 1969; Wets 1988). Many solving programs (CPLEX, MSLip-OSL, OSL-SP, etc.) are cur rently available for this type of model (Dupacova et al. 2000). In this study, we used CPLEX 9.0.

s:t: A1 x1 b1 ; Bk2 x1 Ak2 xk2 bk2 ; k2 [ M2 ; k3 [ M3 ; L {2; 3};

3 4 5 6

Bk2 xapk3 Ak3 xk3 bk3 ; ^ ^k xapkL 2 xapkL 0; L

kL [ ML ;

Objective function The model has multiple objectives. The objective function is dened as the weighted sum of each objective function value in Equation (7): min Obj
2 7 XX t1 i1 T X X

x1 ;xk2 ;xk3 $ 0; M2 {2;:::;N 2 }; M3 {N 2 1;:::;N 3 }

where M2, M3 are the set of indices of divergence arcs at stages 2 and 3, respectively, in the scenario tree; kL is the path (the set of arcs) at the Lth stage in the scenario tree; xkL is the set of decision vectors, such as release and storage of each reservoir at the Lth stage; C kL is a vector of cost coefcients for xkL . In order to relate this conceptual representation to our specic application, the vector of cost coefcients at the Lth stage (C kL ) becomes the composite cost coefcients composed of costs for shortage of water demand, spill, storage and hydroelectric energy objectives. A more detailed description is summarized later in this section; ap(kL)is the immediate ancestor path (set of arcs) of the Lth stage of the kL path, apk3 [ k2 , and apk2 1; pkL are the marginal probabilities of choosing the path in kL P at stage L and pkL 1. The probability ps of realizing specic path S at each stage can be obtained by multiplication of the marginal probability and the conditional probabilities. It is assumed to be 1/n, (Ps 1=nML ), where n(ML) is the number of elements (arcs) of the set ML at the Lth stage. ps is the probability of individual scenarios S and becomes pk3 in this third-stage case; A and B are matrices dening the constraint of the problem; and b represents the right-hand side under path kL. Equation (5) enforces nonanticipativity by requiring that all paths that have the same immediate ancestor maintain the same decision. In ^ Equation (5), xapkL is the immediate ancestor variable that ensures that the planning decision vectors maintain the same decision for all paths having the same immediate ancestor at ^k the Lth stage, and xapkL [ xKL is the set of planning decision L
kL [ML

ci Z t 1 i

10 X X

pv 2

7 X

t3 v[Vt i1 7 X

ci Zt;v 2 i

pv 3

t11 v[Vt i1

ci Zt;v 3 i

where ci is the cost coefcient vector of the ith objective L described in Table 3 at the Lth stage; and Zt;v is the ith i objective value in Table 3 at time period t with scenario v. In our application, stages are dened as prediction lead time
Table 3

Priority of the objectives and basic weights of each objective

Priority Obj. no. number Description of objectives

Basic weights

z1 z2

1 2

Minimize the infeasibility Satisfy the minimum requirement ow Minimize the shortage of water demand Attain the nal target storage Minimize the spill at the basin outlet Maximize the storage Maximize the hydroelectric energy generation.

1011 109

z3

107

z4 z5

4 5

106 104

z6 z7

6 7

103 102

29

Y. Lee et al. | Multistage stochastic linear programming model

Journal of Hydroinformatics | 10.1 | 2008

spans, and the rst-stage decision is made during the time period from t 1 to t 2. And at the rst stage, we have only one scenario. The planning variable at the rst stage is the release decision at each dam, which is made before the inows are revealed, whereas the second-stage decisions are allowed to adapt to the revealed inow except the ones that were determined at the rst stage and set by the nonanticipativity constraint. The objective functions and their priorities and basic weights are summarized in Table 3. Some of the objectives are presented for the purpose of specifying a target for the goal programming. Highest preemptive precedence is granted to minimize the constraint violation for hedging against varying inow scenarios in the stochastic model. The second precedence is given to the minimum requirement ow objective at each control point. The minimum requirement ow is an operational constraint, but we specied it as the goal to be attained by applying an extremely high penalty in order to avoid the infeasibility problem. The shortage of water demand is also minimized by applying a very high weight to the penalty. The daily operation model is supposed to attain monthly storage targets of each reservoir at the end of the month. We assumed that the monthly target storage can be assessed by a stochastic monthly operating model in the higher hierarchy of the decision process. Either two-stage SLP model (Lee et al. 2006) or the sampling SDP model (Kim et al. 2007) based on ESP can be used to provide the monthly storage targets. For successful daily operation, the trade-off analysis between storage maximization and release for hydroelectric energy generation has to be made in multiobjective analysis as demonstrated by Kim et al. (2005). However, such analysis is not the focus of this study. The primary purpose of reservoir operation in Korea is water conservation for safe water supply. Therefore, we tried to minimize the spillage at the basin outlet from the perspective of basin-wide water conservation rather than at each dam site individually (Kim & Park 1998a, b; Kim et al. 2000). We also included the storage maximization objective. Keeping a high water level may increase the future safe water supply capability. To maximize hydroelectric energy generation, an iterative linear approximation technique associated with hydroelectric energy generation is used. The hydroelectric energy is calculated using the equation kWht g Qt Ht 1 hourst. This calculation involves

a nonlinear term, which is described by the product of two variables, head (Ht) and ow (Qt). So the hydroelectric energy function is approximated by kWHt g n o q0 Ht h0 Qt 2 h0 q0 1 hourst as suggested by t t t t Loucks et al. (1981), which is the rst-order linear approximation of the hydroelectric energy function. Starting with the estimated average head, h 0, and ow, q 0, we solve the t t linearized mathematical model iteratively until we nd a convergent optimal pair of head (Ht) and ow (Qt) parameters, which are used to identify the acceptable optimal average ow and head corresponding to the coordinated optimal release and storage levels among the dams in the basin. The results of the hydroelectric energy generation are usually well matched with those in the historical records, if we use appropriate turbine head-ow-efciency curves and assume efcient operation, during the post-analysis phase (Kim 1999).

Model constraints The model is formulated as the multi-stage SLP model based on the multi-period optimization with embedded dynamic network ow. The network ow diagram of the Nakdong River Basin in Korea is represented in Figure 3.

Figure 3

Network ow diagram of Nakdong River Basin in Korea.

30

Y. Lee et al. | Multistage stochastic linear programming model

Journal of Hydroinformatics | 10.1 | 2008

We classied the nodes into eight types which are represented in the legend of Figure 3. Each node has its own characteristics. The sub-basin is the super-source of the network ow. The reservoir storage is described by the ow into carryover nodes, which stores water from one period to the next. The use of carryover nodes therefore means that the network ow can be modelled by the multi-period, dynamic network model. The storage capacity of the reservoir is also divided into multiple zones, each of which is represented by a segment of the carryover nodes corresponding to the storage zones of the reservoir (Kim et al. 2000). The hydroelectric energy is calculated at power plant nodes. These power plants have a maximum ow capacity. Whenever the release of the reservoir is larger than the ow capacity of the power plant and the water surface elevation is higher than the spillway crest, the overow is spilled through the spillway. The release from the reservoir and local ow from the sub-basin are gathered into the control points, where the violation in the minimum required ow and the ow over the ooding level are suppressed using the goal programming technique. The demand sites represent the place or the demand sector for the delivery of water. The consumptive use of the water at the demand site ows to the terminal nodes. The ow from the nal control points FIXED GRAPHICS to the terminal nodes is dened as the excess outow at the basin outlet, and high weighting factors are applied to reduce any such excessive outow from the basin for the purpose of basin-wide water conservation (Kim & Park 1998a, b; Kim et al. 2000) The detailed objective function and constraint equations are presented in the appendix. In this section, the constraints sets are summarized. The constraints are grouped into six sets: (a) Flow conservation constraints. (b) Physical limitation for each node and arc. (c) Goal programming constraints. (d) Hydropower linearization constraints that are the rstorder linear approximation of the hydroelectric energy function. (e) Spillage constraints to prevent spillage unless the water storage level is higher than the spillway crest elevation. Most Korean dams have spillways with radial gates which allow the overow release to be controlled. No

dams are equipped with river outlets or sluiceways, except for hydroelectric and emergency release. Hence, spillage is modelled specically by mixed integer variables which prevent spillage when the storage level is under the spillway crest elevation. (f) Non-anticipativity constraints for the daily release decision at each divergence point in the scenario tree. These constraints have to coincide for the scenarios which have a common path up to each divergence point.

CASE STUDY: NAKDONG RIVER BASIN IN KOREA

We applied the model to the daily operation of the Nakdong River Basin, multi-reservoir system in Korea. Basin outlook The Nakdong River Basin is the second largest watershed in the Korean peninsula, as shown in Figure 4. It occupies approximately 23,817 km2, which constitutes about 24% of the South Korean land area. It is located in the temperate region, with four distinct seasons. The annual average precipitation is about 1,231 mm, two-thirds of which falls between June and September. There are 7 dams (5 multipurpose and 2 water supply) operating in parallel. In this study, we considered the 4 major multipurpose dams, excluding the 3 small dams.

Figure 4

Nakdong River Basin in Korea.

31

Y. Lee et al. | Multistage stochastic linear programming model

Journal of Hydroinformatics | 10.1 | 2008

Scenario generation The streamow scenarios are generated for the daily stochastic model using RRFS with meteorological forecast information. The rst-stage scenarios are generated by RRFS with RDAPS rainfall forecast input data. The second-stage scenarios by RRFS with GDAPS rainfall forecast information are supplemented by the historical maximum and minimum rainfall data, because RRFS may not cover all possible historical extreme situations. The third-stage ESP scenarios generated by RRFS with 20-year daily rainfall data are supplemented by the historical maximum and minimum rainfall. RDAPS and GDAPS data are obtained from KMA and analyzed by the KOWACO Weather Information Service System, the meteorological information data analysis system of KOWACO (KWRA 2003). The accuracy of the generated streamow is represented in Figure 5. Figure 5 shows the mean absolute percent error

(MAPE) of the generated streamow using the historical mean, maximum and minimum rainfall, RDAPS rainfall forecasts and GDAPS rainfall forecasts. MAPE is the average of the absolute values of the relative error rate between generated streamow and observed streamow. It represents forecasting error, regardless of seasonal ow uctuation. The accuracy of the generated streamow is very low due to the uncertainty of the streamow in nature, as shown in Figure 5. However, MAPE of the streamow generated from meteorological forecasts of RDAPS and GDAPS is lower than that of historical rainfall. The streamow generated using the RDAPS rainfall forecasts tends to have the smallest error among all forecasts. However, the forecasting lead-time span of RDAPS information is only 2 days. For the mid-term (up to 10 days), GDAPS information is available. The streamow from GDAPS rainfall forecasts is better than that from the historical average or maximum: however, it is not sufciently accurate from a practical point

Figure 5

MAPE of the generated streamow using RRFS.

32

Y. Lee et al. | Multistage stochastic linear programming model

Journal of Hydroinformatics | 10.1 | 2008

of view. The effect of the meteorological information will be discussed later, based on the comparison of the simulation results with and without meteorological information. Although the streamow generated from historical maximum and minimum rainfall may be inaccurate, it is included in the scenario tree to reduce the risk of serious ooding or drought during the process of choosing the nonanticipating decisions. The added value of updating the non-anticipating decision is discussed in the fourth section. Solution procedure The model was constructed using a Microsoft Visual C 6.0 application with the ILOG Concert technology library 2.0. It comprises over 123,658 columns and 84,129 rows for 100 scenarios. Since the model is a large-scale linear programming model, we could solve the model using the linear programming solver ILOG CPLEX 9.0. As the number of scenarios increases, the model becomes very large, as shown in Tables 4 and 5. In Table 4, when spillage control constraints are excluded, the model was constructed for 60 scenarios (1 scenario for the rst stage, 3 scenarios for the second stage and 20 scenarios for the third stage) and solved in 5 s (19,322 dual simplex iterations, 2 s CPLEX solving time) with a Pentium 4 PC (3.2 GHz CPU and 1GB RAM). Even when the number of scenarios was increased to 300, it was solved in only 3 min 40 s (112,229 dual simplex iterations). When spillage control constraints are included, the model becomes a mixed-integer programming model with a consequent exponential increase in the solution time, as shown in Table 5. For the same 60 scenarios, the model was solved in 8 min
Table 4

(217,182 dual simplex iterations, 469 s CPLEX solving time), which is longer than the case with spillage control constraints excluded. Nevertheless, this 8 min solution time remains in the acceptable range. However, when the number of scenarios was increased to 300, the computation time lengthened to more than 30 h, which is unacceptable for the daily planning model. For more than 60 scenarios, it is recommended to exclude spillage control constraints, even though this can make spill occur when it is physically impossible. For example, when one of the parallel reservoirs has good storage whose water level is lower than the spillway crest elevation, while the other dams storage is insufcient, it may force the water spill downstream in order to meet downstream requirements so as to derive the rst-stage solution, even if water cannot be delivered downstream. The model may derive impracticable solutions in this case, but it remains rare because in practice we use SLP primarily for water supply operation during the non-ood season. To overcome such an impracticable solution, the solution needs to be corrected by a simulation program. For further research, special solution algorithms like the L-shaped method may be required when cases with a large number of scenarios are modelled with spillage control constraints included.

RESULTS AND DISCUSSION

We perform a real-time simulation under uncertainty, in which we assume that historical inow will occur during

Problem size and the solution time without spillage control constraints

Number of scenarios

Columns

Rows

Non-zero elements

Number of iterations

Cplex CPU time (s)

Total time (s)

1 (1 1 1) 9 (1 3 3) 15 (1 3 5) 60 (1 3 20) 100 (1 5 20) 300 (3 5 20)

3,648 34,795 104,011 173,227 346,267 1038,352

3,362 24,474 71,388 118,302 235,586 704,320

8,685 83,907 251,067 418,227 836,127 2507,596

418 3,191 14,949 19,332 38,786 112,229

0.02 0.18 1.52 1.55 5.17 50.92

0.87 1.70 3.94 4.66 13.26 220.09

33

Y. Lee et al. | Multistage stochastic linear programming model

Journal of Hydroinformatics | 10.1 | 2008

Table 5

Problem size and the solution time with spillage control constraints

Number of scenarios

Columns

Rows

Non-zero elements

Number of iterations

Cplex CPU time (s)

Total time (s)

1 (1 1 1) 9 (1 3 3) 15 (1 3 5) 60 (1 3 20) 100 (1 5 20) 300 (3 5 20)

3,798 36,145 106,261 182 227 361,267 1083,352

3,662 27,174 75,888 136,302 265,586 794,320

8,897 85,816 254,249 430,955 857,340 2571,236

543 11,668 86,517 217,182 847,768 4420,930

0.87 17.02 118.27 468.96 1,326.08 110,046.72

1.72 18.68 122.10 477.48 1,338.94 110,260.21

the study period. However, for real-time operation with the stochastic model, the future inow scenarios are generated using RRFS based on the meteorological information. For real-time simulation, the daily release plan for each reservoir is obtained by the stochastic model, and then used as todays release plan in a reservoir simulation model (in the real situation, the use of a multi-reservoir operation simulation run may not be necessary but the actual reservoir operation will be carried out regardless). The multireservoir simulation model may verify the release plan of the stochastic model, and if there is any physical violation, the release plan may be modied to meet the limitation. The actual storage at the end of the day is calculated based on the real inow (the historical inow was presumed as the real inow in the real-time simulation process). A discrepancy between the actual storage attained with the real inow and the desired storage calculated by the stochastic model is unavoidable due to the inow uncertainty. To resolve this discrepancy, the reservoir operation plan from the stochastic model is updated every day with the actual storage data until the end of the study period. Upon implementation of the rst-stage decision on the release of each reservoir with the assumed real inow, the updated actual storage from the reservoir simulation becomes the initial storage for the stochastic model to calculate the release plan for the next day (Kim et al. 2000). We evaluate the added value gained by using a stochastic model in a real-time environment based on the value of information concept (Birge & Louveaux 1997). The expected value of perfect information (EVPI) and the value

of the stochastic solution (VSS) are evaluated. EVPI is the difference between WS (wait-and-see, the solution of perfect information) and the stochastic solution of the recourse problem (RP), which measures the value of perfect information about the future. WS is the expected or average return, in the long run, if we have had perfect information before a decision was made. To calculate this value, we choose the best alternative for each presumed scenario and multiply its payoff with the probability of occurrence of that scenario. As a result, WS is the ideal return we can achieve. RP can be dened as the expected or average return of the nonanticipativity decision with a stochastic model when the assumed scenarios in the stochastic model are reproduced. To evaluate this value, we choose the best decision for hedging against varying inows and calculate the expected value when the best decision is applied to all presumed scenarios. RP is the return attainable in the long run. VSS is the difference between RP and EV (the solution from the expected value problem). It measures the benet of the stochastic model compared to the decision with no information. EV can be dened as the expected value with no information in the long run, in the absence of any information before a decision is made. To calculate this value, we choose the best solution for average inow of the presumed scenarios and calculate the expected value when the solution is applied to all presumed scenario. WS, RP and EV are dened as the expected return of the operation results using the model when the projected scenarios in the stochastic model are reproduced based on

34

Y. Lee et al. | Multistage stochastic linear programming model

Journal of Hydroinformatics | 10.1 | 2008

the value of information concept (Birge & Louveaux 1997). However, it is almost impossible to expect that the inow scenarios assumed will be reproduced exactly in the daily operation. In real-time operation, the release plans are updated every day. Therefore, although the meaning is not exact in the truest sense, we interpreted the expected return in WS, RP and EV as being the results of the real-time operation using the release plan from the model; WS with the perfect rainfall information, RP with the probabilistic forecasting information on the multi-stage, scenario tree and EV with the historical average rainfall information. We compared the real-time simulation results of the stochastic model with those of deterministic models that were divided into two cases: one with no information (use only the average inow information) and the other with perfect information. To evaluate the value of the meteorological forecasts, we performed two separate, real-time simulation studies according to the existence of the forecasted meteorological information. The rst study is for the 20 years from January 1983 to December 2002, during which no forecasted meteorological information is available. We therefore used the historical minimum rainfall data instead of RDAPS and GDAPS forecasts. The second study is for the 3 years from January 2003 to December 2005, in which we assumed that the RDAPS and GDAPS forecast information is available. For all studies, the historical storage at the end of the month is assumed as the monthly storage targets of each reservoir, to evaluate the pure effect of the daily stochastic planning model in the real-time simulation.

The results of real-time simulation without meteorological information We performed the real-time simulation for the 20 years from January 1983 to December 2002 without meteorological information in order to investigate the effect of the stochastic model with RRFS forecasting. The results are summarized in Table 6, which shows the VSS results when we apply the daily SLP model to the Nakdong River Basin multi-reservoir system in Korea in the real-time operational environments. EV, RP and WS are the expectations of the operational results of using the model. The stochastic solution (RP) reduced the water supply decit and the spill amount by 21 MCM and 905 MCM, respectively, compared to the case of using average inow, which is shown in the VSS column. The RP rate shows the incremental performance of RP relative to the case without any information. It is calculated as the ratio of the incremental performance of EV to the difference between EV and WS. Average storage of the RP rate reaches 43% of the perfect information (WS) results. A box plot of the results of the real-time simulation for 20 years is depicted in Figure 6. The variance of EV average storage is larger than that of RP and WS, which indicates that EV is more inuenced by the future inow uncertainty. The spill results conrm that WS and RP can cope with the ooding condition of the high ow scenario effectively by reserving appropriate ood control capacity of the reservoir in advance. The amount of hydroelectric energy generation of RP and WS is also larger than that of the EV case for all inow scenarios.

Table 6

Results of the real-time simulation under uncertainty without meteorological information from January 1983 to December, 2002 (units: MCM) RP ratep

Objectives

EV

RP

WS

EVPI(WS vs. RP)

VSS(RP vs. EV)

Deviation from the nal storage target Water supply decitpp Amount of spillpp Average storageppp Hydroelectric energy(GWh)pp

36 21 1,542 1,403 515

346 1,433 542

4 291 17 11

32 21 905 13 16

89% 100% 76% 43% 59%

637 1,416 531

p RP rate (RP EV)/(WS EV), pp yearly total in the basin, ppp daily total in the basin

35

Y. Lee et al. | Multistage stochastic linear programming model

Journal of Hydroinformatics | 10.1 | 2008

Figure 6

Box plot of the results of the real-time simulation under uncertainty.

This result demonstrates the effect of the nonanticipativity decision of the stochastic model and the added information gained from using the RRFS forecasting ability while preserving the serial correlation. The added value of the meteorological information with SLP In order to examine the effect of the meteorological information on the SLP results, a real-time simulation is performed for the 3 years from January 2003 to December 2005 and the results are summarized in Table 7. Table 8 presents a comparison of the results with and without meteorological information. As shown in Table 7, the RP results reduce the water supply decit and the spill amount by 3 MCM and 909
Table 7

MCM, respectively, compared to EV, as shown in the VSS column. The performance increase of the RP results on the spill amount accounts for up to 93% of the WS results, conrming the effectiveness of using the stochastic model (RP) in real-time operation to avoid spill occurrences. The results in the other objective categories also demonstrate the usefulness of the stochastic model under uncertainty. As shown in Table 8, the comparison of performance with and without meteorological information conrms the improved result obtained with the added value of the meteorological information. The spill amount is decreased by 293 MCM (30% of the RP rate) using the improved streamow forecast generated with the meteorological information. The average storage and hydroelectric energy generation are

Results of the real-time simulation under uncertainty with meteorological information from 2003 to 2005 (units: MCM)

Objectives

EV

RP

WS

EVPI (WS vs. RP)

VSS (RP vs. EV)

RP rate p

Deviation from the nal storage target Water supply decitpp Amount of spillpp Average storagepp p Hydroelectric energy(GWh)pp

18 3 1,633 1,423 517

1 724 1,431 541

657 1,441 548

1 67 10 7

17 3 909 8 24

94% 100% 93% 44% 77%

RP rate (RP EV)/(WS EV),

pp

yearly total in the basin,

p pp

daily total in the basin

36

Y. Lee et al. | Multistage stochastic linear programming model

Journal of Hydroinformatics | 10.1 | 2008

Table 8

The comparison of results with and without meteorological information from 2003 to 2005 (units: MCM)

Without meteorological information Objectives RP RP rate

With meteorological information RP RP rate Improvements RP RP rate

Deviation from the nal storage target Water supply decitp p Amount of spillpp Average storageppp Hydroelectric energy(GWh)p p

94%

94%

0%

1,017 1,430 531

100% 63% 39% 45%

724 1,431 541

100% 93% 44% 77%

2 293 1 10

0% 30% 6% 32%

RP rate (RP EV)/(WS EV),

pp

yearly total in the basin,

ppp

daily total in the basin

increased by 1 MCM and 10 GWh, respectively, compared to the case without meteorological information. In addition, we evaluated the cases in which only one meteorological forecast is availableL RDAPS or GDAPS. In the RDAPS case (i.e. employing RDAPS forecasting information at the rst stage) we applied the historical average rainfall instead of GDAPS forecasts at the second stage because the forecasting lead-time of RDAPS is only 2 days. In the GDAPS case, we used GDAPS information all the way from the rst stage to the second stage without RDAPS information. The results are summarized in Table 9. It shows that the use of only RDAPS information or only GDAPS information is more efcient than that of using no meteorological information. The results of using only GDAPS information decreased spill by 30 MCM and increased hydroelectric energy generation by 1 GWh, respectively, compared to the case without meteorological information. The results of using only RDAPS information also decreased spill by 269 MCM and increased hydroelectric energy generation by 8 GWh compared to the case without meteorological information. When we compared the results of using only RDAPS information and that of using only GDAPS information, the RDAPS case is more efcient than the GDAPS case because the accuracy of RDAPS is more reliable than GDAPS, even if GDAPS has a longer forecasting lead-time span. Moreover, the accuracy

of the short-term forecasts may affect the results more because we apply only a rst day release decision in the real-time simulation process and update the decision every day. Finally, the subsequent application of GDAPS information to the RDAPS result, as originally suggested in the model, produces the most efcient results.

Table 9

The comparison of results for each case of using meteorological information from 2003 to 2005 (units: MCM)

Without meteorological Objectives information

With only GDAPS information

With only RDAPS information

With RDAPS and GDAPS information

Deviation from the nal storage target Water supply decitp Amount of spillp Average storagepp Hydroelectric energy(GWh)p

1,017

987

748

724

1,430

1,430

1,430

1,431

531

532

539

541

Yearly total in the basin,

pp

daily total in the basin

37

Y. Lee et al. | Multistage stochastic linear programming model

Journal of Hydroinformatics | 10.1 | 2008

CONCLUSION
In this study, a multi-period, multi-stage SLP with recourse model is developed to derive the daily release plan of each reservoir in a coordinated multi-reservoir system. Stages are dened as prediction lead-time spans. The multi-stage scenarios for the stochastic model are formulated considering the reliability of rainfall radar prediction for a specied forecasting lead-time. The future inows are generated by RRFS based on the meteorological rainfall forecasts. The forecasting model for each stage is selected based on the viability and accuracy of the prediction. For short-term (2 days) rainfall forecasting, RDAPS is applied. For mid-term (up to 10 days) forecasting, precipitation from GDAPS is used as an input for a rainfall runoff model in the ESP procedures. After the tenth day (third stage), the daily historical ensemble rainfall data are applied. The model is applied to a simulation of the daily reservoir operation of the Nakdong River Basin multireservoir system in Korea, under a real-time operational environment. The expected benet of the stochastic model is analyzed quantitatively based on Birge & Louveaux (1997) value of information measure. We also evaluate the added value gained by using meteorological information. We perform two separate, real-time simulation studies according to the existence of the forecasted meteorological information. The rst study is for the 20 years from January 1983 to December 2002, during which no forecasted meteorological information is available and we therefore use minimum rainfall data instead of RDAPS and GDAPS forecasts. The stochastic solution (RP) reduces the water supply decit and the spill amount, and increases the average storage and hydropower generation compared to the case of using average inow. This result demonstrates the effect of the nonanticipativity decision of the stochastic model and the added information gained from utilizing the RRFS forecasting ability without meteorological forecast information, while preserving the serial correlation. The second study is for the 3 years from January 2003 to December 2005, during which we assume the availability of the RDAPS and GDAPS rainfall forecasting information. From this result, we could estimate the added value gained from using the meteorological forecasting information. The meteorological forecasts obtained with SLP increase the

hydroelectric energy generation while reducing the spill, conrming the effectiveness of using the stochastic model in real-time operation with meteorological forecasts in the presence of uncertainty. In this study, the historical storage at the end of the month was used as the monthly storage target for each reservoir, to evaluate the pure effect of the daily stochastic planning model in a real-time simulation. The monthly stochastic model is required for long-term storage planning, while the model we developed here is capable of short-term real-time operation. Special solution algorithms like the L-shaped method may be required when cases with a large number of scenarios are modelled with spillage control constraints included.

ACKNOWLEDGEMENTS
This research was supported by a grant (code 1-6-2) from the Sustainable Water Resources Research Center of 21st Century Frontier Research Program and partly by the Brain Korea 21 Project in 2007.

REFERENCES
Beale, E. M. L. 1955 On minimizing a convex function subject to linear inequalities. J.R. Statist. Soc. Ser. B 17 (2), 173 184. Birge, J. R. & Louveaux, F. 1997 Introduction to Stochastic Programming. Springer-Verlag, Berlin. Can, E. K. & Houck, M. H. 1984 Real time reservoir operating by goal programming. J. Wat. Res. Plann. Mngmnt. 110 (3), 297309. Dantzig, G. B. 1955 Linear programming under uncertainty. Mngmnt Sci. 1 (3), 197 206. Dantzig, G. B. & Madansky, A. 1961 On the solution of two-stage linear programming under uncertainty. In Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and Probability. University of California Press. Berkeley, CA. Day, G. N. 1985 Extended streamow forecasting using NWSRFS. J. Wat. Res. Plann. Mngmnt. 111, 147170. Dupacova, J., Consigli, G. & Wallace, S. W. 2000 Scenarios for multistage stochastic programs. Annals. Oper. Res. 100, 25 53. Eschenbach, E. A., Magee, T., Zagona, E., Gorano, M. & Shane, R. 2001 Multiobjective operations of reservoir systems via goal programming. J. Wat. Res. Plann. Mngmnt. 127 (2), 108120. Houck, M. H. 1979 A chance constrained optimization model for reservoir design and operation. Wat. Res. Res. 15 (5), 1011 1015.

38

Y. Lee et al. | Multistage stochastic linear programming model

Journal of Hydroinformatics | 10.1 | 2008

Jacobs J., Freeman, G., Grygier, J., Morton, D., Schultz, G., Staschus, K. & Stedinger, J. 1995 Stochastic optimal coordination of river-basin and thermal electric systems (SOCRATES): a system for scheduling hydroelectric generation under uncertainty. Annals Oper. Res. 59, 99 133. Kim, S. K. 1999 Hydro energy from multiple reservoir operation. In HYDROS FUTURE Technology, Markets, and Policy, Proceedings of the WATERPOWER99 Conference July 6-9, 1999, MGM Grand Hotel and MGM Grand Conference Center, Las Vegas, NV (ed. P. A. Brookshier). Water Resources Session 23: Decision Support Systems II, No 2. Kim, S. K., Lee, Y. D., Kim, J. H. & Ko, I. H. 2005 A multiple objective mathematical model for daily coordinated multi-reservoir operation. Wat. Sci. Technol. Wat. Supply 5 (3-4), 8188. Kim, S. K. & Park, Y. J. 1998a Coordinated multireservoir operating model. In May 26 29 1998 Spring Meeting, American Geophysical Union, Boston, MA. Kim, S. K. & Park, Y. J. 1998b A mathematical model for coordinated multi-reservoir operation. J. Korea Wat. Res. Assoc. 31 (6), 779 793 (in Korean). Kim, S. K., Park, Y. J. & Kim, J. H. 2000 Simulation of the real-time multiple reservoir operating environment with a multireservoir operation optimization model. In ASCEs 2000 Joint Conference on Water Resources Engineering and Water Resources Planning and Management. Kim, Y. O., Eum, H. -I., Lee, E. G. & Ko, I. H. 2007 Optimizing operational policies of a Korean multi-reservoir system using sampling stochastic dynamic programming with ensemble streamow prediction. J. Wat. Res. Plann. Mngmnt 133 (1), 414. Kracman, D. R., McKinney, D. C., Watkins, D. W. Jr & Lasdon, L. S. 2006 Stochastic optimization of the Highland Lakes System in Texas. J. Wat. Res. Plann. Mngmnt. 132 (2), 6270. KWRA (Korea Water Resources Association) 2003 Study on Watershed-based Seasonal Forecast Rainfall for Dam Operation Application. Korea Water Resources Corporation (in Korean). Daejeon. Labadie, John W. 2004 Optimal operation of multireservoir systems: state-of-art review. J. Wat. Res. Plann. Mngmnt. 130 (2), 93111. Lee, Y. D., Kim, S. K. & Ko, I. H. 2006 Two-stage stochastic linear programming model for coordinated multi-reservoir operation. In The 2006 Operations Management Conference about Operating Reservoirs in Changing Conditions sponsored by EWRI of ASCE, August 14-16, 2006, Sacramento, CA. Loucks, D. P., Stedinger, J. R. & Haith, D. A. 1981 Water Resources Systems Planning and Analysis. Prentice-Hall. Englewood Cliffs, NJ.

Mujumdar, P. P. & Ramesh, T. S. V. 1997 Real-time reservoir operation for irrigation. Wat. Res. Res. 33 (5), 1157 1164. Pereira, M. V. F. & Pinto, L. M. V. G. 1985 Stochastic optimization of a multireservoir hydroelectric system: a decomposition approach. Wat. Res. Res. 21 (6), 779 792. Ponnambalam, K., Vannelli, A. & Unny, T. E. 1989 An application of Karmarkars interior-point linear programming algorithm for multireservoir operations optimization. Stochast. Hydro. Hydraul. 3, 17 29. Prekopa, A. 1995 Stochastic Programming. Kluwer, Dordrecht. ReVelle, C., Joeres, E. & Kirby, W. 1969 Linear decision rule in reservoir management and design. 1: Development of the stochastic model. Wat. Res. Res. 5 (4), 767 777. Sei, A. & Hipel, K. W. 2001 Interior-point method for reservoir operation with stochastic inows. J. Wat. Res. Plann. Mngmnt. 127 (1), 48 57. Sen, S. & Higle, J. L. 1999 An introductory tutorial on stochastic linear programming models. Interfaces 29 (2), 33 61. Stedinger, J. R., Sule, B. F. & Loucks, D. P. 1984 Stochastic dynamic programming models for reservoir operation optimization. Wat. Res. Res. 20 (11), 1499 1505. Turgeon, A. 2005 Daily operation of reservoir subject to yearly probabilistic constraints. J. Wat. Res. Plann. Mngmnt. 131 (5), 342 350. US Army Corps of Engineers 2006 Streamow Synthesis and Reservoir Regulation Model. Available at: http://www.nwd-wc. usace.army.mil/report/ssarr.htm Van Slyke, R. & Wets, R. 1969 L-shaped linear programs with application to optimal control and stochastic programming. SIAM J. Appl. Math. 17, 638663. Watkins, D. W., McKinney, D. C., Lasdon, L. S., Nielsen, S. S. & Martin, Q. W. 2000 A scenario-based stochastic programming model for water supplies from the Highland Lakes. Int. Trans. Oper. Res. 7 (3), 211 230. Wets, R. J.-B. 1988 Large scale linear programming techniques. In Numerical Techniques for Stochastic Optimization (ed. Y. S. Ermoliev & R. J. -B. Wets), Springer, New York. pp. 65 93. Wurbs, R. A. 1993 Reservoir-system simulation and optimization models. J. Wat. Res. Plann. Mngmnt. 119 (4), 455 472. Yakowitz, S. 1982 Dynamic programming applications in water resources. Wat. Res. Res. 18 (4), 673 696. Yazicigil, H., Houck, M. H. & Toebes, G. H. 1983 Daily operation of a multipurpose reservoir system. Wat. Res. Res. 19 (1), 111. Yeh, W. W.-G. 1985 Reservoir management and operations models: a state-of-the-art review. Wat. Res. Res. 21 (12), 1797 1818.

First received 3 July 2006; accepted in revised form 20 August 2007

39

Y. Lee et al. | Multistage stochastic linear programming model

Journal of Hydroinformatics | 10.1 | 2008

APPENDIX (MATHEMATICAL MODEL)

Denition of sets CLu: CUl: CP: DS: N: PS: RE: SB: SP: TN: TP: V(t): Set of lower nodes connected to an upper node u in the basin network Set of upper nodes connected to a lower node l in the basin network Set of control points Set of demand sites Set of nodes Set of power stations Set of reservoirs Set of sub-basins Set of spillways Set of terminals (sink node) Set of time periods Set of scenarios (arc in the scenario tree) at time period t. Denition of variables Dt;v : d Et;v : p Decit of the water demand at demand site d [ DS, at time period t [ TP with scenario v Hydroelectric energy generation of powerplant p [ PS, at time period t with scenario v
t; ET nv : Surplus to target of node n [ N, at time period t

aa(v [ V(t)) [ V(t 2 1): The immediate ancestor arc of scenario v at time period t. It is required to keep track of the carryover storage. apv [ Vt: The immediate ancestor path (set of arcs) of scenario v at the stage belonging to the time period t. It is dened for the nonanticipativity constraints because the planning decision (the daily release plan) has to coincide for the scenarios which have a common path up to each divergence point ci L : The cost coefcient vector of the ith objective at the Lth stage dmdt : d eht;v : p eqt;v : p init_sr : lbt : n mrf t : c pv : L rfrd : rot;v : b The water demand of demand site d [ DS, at time period t The estimated head of powerplant p [ PS, at time period t with scenario v The estimated ow of powerplant p [ PS, at time period t with scenario v The initial storage of reservoir r [ RE The lower bound of node n [ N, at time period t The minimum requirement ow of control point c [ CP, at time period t The marginal probabilities of choosing scenario v at stage L The return ow rate from demand site d [ DS, at time period t The runoff to a sub-basin, b [ SB, at time period t with scenario v spillable_strs: The reservoir storage at the spillway crest elevation of spillway s [ SP tailwaterp: ubt : n The tailwater elevation of powerplant p [ PS The upper bound of node n [ N at time period t

with scenario v
t; ET nv2 : Decit from target of node n [ N, at time period t

with scenario v Qt;v : u;l Flow from node u (upper node) to node l (lower node) at time period t with scenario v Rt;apv : Release of reservoir r [ RE, at time period t having r apv [ Vt as the ancestor path (planning decision variable) St;v : r Storage of reservoir r [ RE, at time period t with scenario v
t; V r v and t; V r v2 : Violation of the control constrains at

reservoir r at time period t with scenario v Zt;v : i ith objective value at time period t [ TP with scenario v

Geht;v ; eqt;v : The hydroelectric energy generation p p

coefcient of efciency hill diagram 9.8 24 efciency of estimated head at estimated ow of powerplant p [ PS at time period t

Denition of constants at , r bt : r The coefcients of linearized storage-elevation function of reservoir r [ RE

40

Y. Lee et al. | Multistage stochastic linear programming model

Journal of Hydroinformatics | 10.1 | 2008

Objectives min Obj

2 7 XX t1 i1

Reservoirs ci Z t 1 i
10 X X

pv 2

7 X

t3 v[Vt i1

ci Zt;v 2 i

St;v r

l[CLr

X t;v t; t; Qr;l 2 V r v V r v2 X Qt;v for r [ RE; u;r A10

T 7 X X v X i t;v p3 c 3 Z i t11 v[Vt i1

A1

St21;aav r

u[CU r

v [ Vt;
Minimize the infeasibility Zt;v 1 X t; t; {V r v V r v2 }
r[RE

t [ TP for r [ RE A11

S0 init_sr r A2

Demand sites X A3
t; Qt;v ET dv2 dmdt d u;d

Satisfy the minimum requirement ow Zt;v 2 X t;v ET c

for d [ DS; A12

u[CU d

c[CP

v [ Vt;
r f rd X

t [ TP for d [ DS; c [ CLd > CP; A13 t [ TP

Minimize the shortage of water demand Zt;v 3 X t; ET dv2

d[DS

u[CUd

Qt;v Qt;v u;d d;c

A4

v [ Vt;

Attain the nal target storage ZT;v 4 {ET T ;v r r[RE X ET T;v2 } r A5

For other nodes (CP, PS, OL, SP) X Qt;v u;n X Qt;v n;l A14

u[CU n

l[CLn

for n [ CP < PS < SP; v [ Vt; t [ TP Minimize the spill at the basin outlet Zt;v 5 X X Qt;v u;s A6 (b) Physical limitation for each node and arc Bounds on the storage Maximize the storage Zt;v 6 X 2 St;v r
r[RE

s[SP u[CU s

lbt # St;v # ubt r r r A7 t [ TP

for r [ RE;

v [ Vt;
A15

Maximize the hydroelectric energy generation Zt;v 7 X 2 Et;v p

p[PS

For other nodes (CP, PS, OL, SP) A8 lbt # n X t;v Qu;n # ubt n t [ TP for n [ CP < PS < SP; A16

u[CU n

v [ Vt;
Constraints

(c) Goal programming constraints (a) Flow conservation constraints Attaining the target storage Sub-basins X Qt;v rot;v b;l b for b [ SB;
T; ST;v 2 ET T;v ET r v2 tar_sr r r

v [ Vt;

t [ TP

A9

for

r [ RE; A17

t[CLb

v [ Vt

41

Y. Lee et al. | Multistage stochastic linear programming model

Journal of Hydroinformatics | 10.1 | 2008

Satisfy the minimum requirement ow X

t; Qt;v ET c v2 $mrf t for c[CP; v [ Vt; t[TP u;n c

(e) Spillage control constraints A18 Qt;v # BSt;v ubt s r;s s for s [ SP; r [ CU s > RS; A21

u[CU c

v [ Vt;

t [ TP

(d) Hydropower linearization constraints for each powerstation which are the rst-order linear approximation of the hydroelectric energy function Et;v Geht;v ;eqt;v {eht;v Qt;v eqt;v Ht;v 2 eht;v eqt;v } p p p p r;p p p p p for p [ PS; v [ Vt; t [ TP   Ht;v at St;v bt 2 tailwaterp r r r p r [ RS > CU p ; for p [ PS; A20 A19

BSt;v # St;v =spillable_strs s r r [ CUs > RS;

for s [ SP; A22 t [ TP

v [ Vt;

v [ Vt;

t [ TP

(f) Non-anticipativity constraints on release decision X Rt;apv Qt;v for r [ RE; v [ Vt; t [ TP A23 r r;l
l[CLr