You are on page 1of 12

HYDROLOGICAL PROCESSES Hydrol. Process. 22, 842 853 (2008) Published online 27 June 2007 in Wiley InterScience (www.interscience.wiley.

com) DOI: 10.1002/hyp.6651

Optimal operation of reservoirs for downstream water quality control using linked simulation optimization
Anirban Dhar and Bithin Datta*
Department of Civil Engineering, Indian Institute of Technology Kanpur, Kanpur-208016, India

Abstract:
A methodology is developed for optimal operation of reservoirs to control water quality requirements at downstream locations. The physicochemical processes involved are incorporated using a numerical simulation model. This simulation model is then linked externally with an optimization algorithm. This linked simulationoptimization-based methodology is used to obtain optimal reservoir operation policy. An elitist genetic algorithm is used as the optimization algorithm. This elitist-geneticalgorithm-based linked simulationoptimization model is capable of evolving short-term optimal operation strategies for controlling water quality downstream of a reservoir. The performance of the methodology developed is evaluated for an illustrative example problem. Different plausible scenarios of management are considered. The operation policies obtained are tested by simulating the resulting pollutant concentrations downstream of the reservoir. These performance evaluations consider various scenarios of inow, permissible concentration limits, and a number of management periods. These evaluations establish the potential applicability of the developed methodology for optimal control of water quality downstream of a reservoir. Copyright 2007 John Wiley & Sons, Ltd.
KEY WORDS

reservoir operation; water quality management; linked simulation optimization; genetic algorithm

Received 4 April 2006; Accepted 23 November 2006

INTRODUCTION Water quality management has received increasing attention in the recent past. Poor understandings of aquatic and riparian ecosystems and the lack of quantication of social and environmental objectives are some of the possible reasons for inadequate development of operational models emphasizing water quality aspects. Maintaining water quality downstream of reservoir systems is also becoming an important objective in optimal operation of reservoir systems. Especially in dry periods, maintenance of water quality downstream of a reservoir may become the main objective of operation, while trying to maintain a target storage level. However, owing to various complexities involved, this objective is often ignored. Any meaningful optimal operation strategy should incorporate water quality aspects, otherwise undesirable environmental or ecological problems may arise, e.g. high pollutant concentration, eutrophication, turbidity, etc. A linked simulation optimization-based methodology is developed for optimal operation of reservoir to control downstream water quality. Numerical simulation models are used for simulating the physical processes governing water quality in water bodies such as lakes and rivers. Simulation models are not efcient in prescribing optimal strategies. Simulationresults provide a detailed description of how systems respond to, or are affected by, planning and design solutions, or sets of solutions. Prescription of
* Correspondence to: Bithin Datta, Department of Civil Engineering, I.I.T. Kanpur, Kanpur-208016, India. E-mail: bithin@iitk.ac.in Copyright 2007 John Wiley & Sons, Ltd.

optimal strategies requires formulation and solution of optimization models. Especially in surface water quality management problems, it becomes necessary to evolve meaningful management strategies based on the simulation of the ow and transport processes. These processes are generally very complex and non-linear in nature. Therefore, incorporation of the physical processes, i.e. ows and transport, within an optimization model may be difcult due the size and non-linearity of the problem. One possible remedy is to link the simulation model externally with the optimization algorithm to evolve optimum management strategies for large-scale and real-life problems. In a linked optimization simulation model, the optimization algorithm may obtain gradient information at every stage of iteration by solving the external simulation model. Use of conventional optimization techniques for this purpose becomes difcult, as the Jacobian matrix (Carmichael and Strzepek, 2000) needs to be generated at every stage of iteration. Genetic algorithms (GAs) are based on the repeated evaluation of the tness function for the population at every generation. As a result, linking of the simulation model with a GA-based optimization model becomes computationally simple and efcient. Therefore, a GA-based linked simulationoptimization model is developed for optimal control of water quality downstream of a reservoir. In a riverreservoir system, the water quality management may consist of (a) controlling the reservoir release concentration, (b) controlling the tributary inow concentration, (c) controlling the tributary inow during the

OPTIMAL RESERVOIR OPERATION FOR DOWNSTREAM WATER QUALITY CONTROL

843

management period, (d) controlling the reservoir release during the management period, and (e) increasing the in-stream degradation rate of the pollutant. This paper focuses on the fourth approach, i.e. control of reservoir releases that inuence the mixing process. Optimal strategies for release of water from storage, given downstream water quality requirements, are obtained using a linked simulation optimization model. The reservoir operation policy determines the release from the reservoir in a given management period. This release would control the downstream water quality. The simulation model CE-QUAL-W2 (Cole and Wells, 2003) used is capable of incorporating various physicochemical process, and can simulate the constituents like bacteria, ammonium, nitratenitrite, dissolved silica, iron, chemical oxygen demand, algae, alkalinity, etc. Performance of the methodology developed is evaluated for different operational scenarios for an illustrative system. These performance evaluations are useful in demonstrating the potential applicability of the developed methodology for optimal management of reservoir systems, with downstream water quality control as the main criterion of operation. Many researchers have developed classical optimization-based models for control of water quality in reservoir and river systems. Jaworski et al. (1970) and Loucks and Jacoby (1972) developed optimal strategies for downstream dissolved oxygen and biological oxygen demand control. Reliability-based multipurpose operation of reservoirs for water quality control was proposed by Orlob and Simonovic (1982). Simonovic and Orlob (1982) further extended this approach to consider the stochastic nature of stream ows. Ikebuchi et al. (1982) used dynamic programming to identify optimal releases for turbidity control. Simonovic and Orlob (1984) applied risk-reliability programming for water quality control. Kojiri (1987) extended the work of Ikebuchi et al. (1982) to obtain multi-objective tradeoffs between low ow and turbidity. Dandy and Crawley (1992) solved a linear management model based on a simplied simulation model. Nandalal and Bogardi (1995) used the simulation model of Dandy and Crawley (1992), along with incremental dynamic programming, to control the water quality. Datta and Soraganvi (1996) proposed a two-reservoir multiobjective formulation for hydropower generation and water quality control downstream of a reservoir system. A few simulation-model-based studies are reported in Wu et al. (1996), Willey et al. (1996), Gunduz et al. (1998) and Hejzlar et al. (2002). Ko (1997) integrated the water quality simulation model QUAL2E-UNCAS with CSDUP (a dynamic programming solver) to evaluate management strategies considering both water quality and quantity. Hayes et al. (1998) presented a method for enhancing water quality in hydropower operation by considering bounds in the water quality indicators. Carmichael and Strzepek (2000) linked the water quality simulation model QUAL2E with the non-linear optimization system MINOS to obtain optimal management strategies for waste treatment cost minimization. Dai and
Copyright 2007 John Wiley & Sons, Ltd.

Labadie (2001) linked a FrankWolfe non-linear programming algorithm with QUAL2E and a river basin network ow model (MODSIM) for management of water quality. Chaves and Kojiri (2003) linked an articial neural network (ANN) approximator as a water quality simulation model with dynamic programming to satisfy fuzzy water quality objectives. Chaves et al. (2003) linked a one-dimensional water quality model with a dynamicprogramming-based optimization model. Comprehensive reviews by Yeh (1985), Wurbs (1993), Labadie (1997) and the most recent one by Labadie (2004) reveal that GAs so far have had very limited application in riverreservoir system management. Esat and Hall (1994) and Fahmy et al. (1994) applied GAs for reservoir system management and compared the potential applicability of GAs with dynamic programming. Olivera and Loucks (1997) dened GA-based multiple reservoir operating policies using system rule curves and an individual storage target balancing function. Chang and Chen (1998) used both binary-coded and real-coded GAs for optimization of a ood control reservoir. Wardlaw and Sharif (1999) and Sharif and Wardlaw (2000) found GAs to be suitable for solving non-linear objective functions, and for generating solutions close to optimum. Cai et al. (2001) demonstrated the applicability of GAs to solving large-scale non-linear water management problems covering multiple time periods. Huang et al. (2002) applied GA-based stochastic dynamic programming to deal with the dimensionality problem associated with multiplereservoir systems. Nicklow et al. (2003) used a GA for controlling channel bed morphology in operation of multi-reservoir systems. Ndiritu (2003) applied a multipopulation GA for optimizing a system of two reservoirs that supplies monthly varying demands and environmental ow requirements. Chen (2003) used a real-coded GA to optimize the operating rule curves dening the storage levels and target release. Celeste et al. (2004) demonstrated the advantages of using GAs for real-time reservoir operation. In the present study, a GA-based linked simulationoptimization methodology is developed using an elitist GA for optimal control of water quality downstream of a reservoir while minimizing deviations from a prescribed storage target. The numerical water quality model considers two-dimensional pollutant transport, and the operation policy is applicable for short-term operation. The performance of the methodology is evaluated for an illustrative study area. In the evaluations reported herein, the inows to the reservoirs are assumed to be deterministic with no forecasting errors. WATER QUALITY SIMULATION MODEL In the proposed methodology, the optimal operation policy is based on the simulation of the pollutant transport processes. A numerical simulation model is externally linked to the optimization model for this purpose. Water quality simulation models describe the functional relationship between different physical, biological and chemical processes that are assumed to occur in water body
Hydrol. Process. 22, 842 853 (2008) DOI: 10.1002/hyp

844

A. DHAR AND B. DATTA

(Ko, 1997). These functional relationships are based on conservation of mass, energy, and momentum principles. In their most general form, these governing equations are systems of partial differential equations with sources/sink terms representing possible decision or control alternatives. Numerical or analytical solutions of these equations are carried out by the simulation models. Large numbers of models are available to simulate water quality in river or riverreservoir systems (Donigian et al., 1995; Rajbhandari, 1995). Only a few models have included hydrodynamics within their algorithm structures and have the capability of simulation for transient conditions. The simulation model utilized in this study is the surface water ow and transport simulation model developed by the US Army Corps of Engineers, CEQUAL-W2 (Cole and Wells, 2003). CE-QUAL-W2 is a two-dimensional (longitudinal/vertical) laterally averaged, coupled model of hydrodynamics and water quality. It is primarily suited for relatively long and narrow water bodies exhibiting water quality gradients in the longitudinal and vertical axes. The model predicts water surface, velocities, temperatures and constituent concentrations. Any combination of constituents can be included or excluded for the simulation. The model requires (a) geometric data, (b) initial condition, (c) boundary condition, (d) hydraulic parameters, (e) kinetic parameters, and (f) calibration data. CE-QUAL-W2 solves the nite-difference form of the laterally averaged equations of uid motion and transport. The governing equations are presented below (Cole and Wells, 2003). Horizontal momentum equation: UB UUB WUB C C D t x z g cos B gB sin C g cos B x 1 B xx 1 B xz C C C qBUx x z

where is the laterally averaged constituent concentration, Dx is the longitudinally averaged temperature and constituent dispersion coefcient, Dy is the vertical temperature and constituent dispersion coefcient, q is the lateral inow or outow mass ow rate of constituent per unit volume, and S is the laterally averaged source/sink term. Free water surface elevation: B D t x
h h

UB dz

qB dz

where B is the channel width at water surface. Continuity equation: UB WB C D qB x z 5

Equation of state, relating density as function of temperature and solids concentration: D f Tw , TDS , ISS 6

where Tw is the temperature, TDS is the concentration of total dissolved solids, and ISS is the concentration of inorganic suspended solids.

OPTIMIZATION MODEL The optimal reservoir operation policy is obtained as a solution of an optimization model. The aim of the reservoir operation is to maintain or control water quality downstream of the reservoir. In general, water quality is more important during dry periods. Therefore, the explicit objective of operation is to minimize the deviation between prespecied target storage and the ending storage for each of the management periods. The specied target storage level can be considered as a surrogate economic objective. In real-life cases, a target storage level would depend upon various objectives of operation. The optimization model uses the simulation results from the externally linked simulation model to determine optimal management strategies. Therefore, the size of the optimization model in terms of number of decision variables would be much smaller, compared with one that uses an embedding technique. The single objective management model formulated in this study considers the single objective of minimizing the summation of normalized squared deviation of actual storage from specied target storage at the end of each management period, subject to the constraint of maintaining water quality standards at downstream control points. This constraint is incorporated through an externally linked numerical simulation model. The decision variables of the management model are the temporal reservoir release quantities. This transient downstream water quality management model can be mathematically represented as Minimize
8t2T

dz x 1

where U is the horizontal velocity, W is the vertical velocity, B is the channel width, x is the x-direction lateral average shear stress, y is the y-direction lateral average shear stress, is the density, is the channel inclination, g is the acceleration due to gravity, q is the lateral inow per unit width, and is the water surface. Vertical momentum equation: 1 P D g cos z where P is the pressure. Constituent transport equation: B UB WB C C t x z x BDz z D q B C S B z
Copyright 2007 John Wiley & Sons, Ltd.

BDx

St

Starget Starget

Hydrol. Process. 22, 842853 (2008) DOI: 10.1002/hyp

OPTIMAL RESERVOIR OPERATION FOR DOWNSTREAM WATER QUALITY CONTROL

845

where Starget is the target storage, St is the storage at the end of the management period t, and T is the time horizon considered, subject to: Mass balance constraint. Continuity of reservoir storage between discrete time intervals must be satised while accounting for the accumulation or depletion of the reservoir storage over the management time horizon. These constraints are used to dene a feasible solution space to ensure feasible reservoir operations: St D St
1

A suitable non-linear optimization technique is required to solve this optimization model for reservoir management to control downstream water quality. An elitist GA is utilized to solve the single-objective non-linear optimization model. SOLUTION METHODOLOGY The methodology developed is presented schematically in Figure 1. The simulation model used in this study is the CE-QUAL-W2 numerical hydrodynamic water quality simulation model. This simulation model is linked to the GA-based optimization model. In the linked simulationoptimization approach, the water quality simulation constraint, Equation (9), is actually satised by solving the numerical water quality simulation model (CE-QUAL-W2) externally linked to the optimization model. The implementation of the GA is described here. In binary-coded GAs (Holland, 1975; Goldberg, 1989) the object variables, i.e. release values in different management periods, are encoded in nite-length binary strings D R1 , R2 , R3 , . . . , Rt , . . . . A linear decoding R function is used to decode the object variable. It linearly maps the decoded value of the binary string in the desired interval [Rt,min , Rt,max ]. In binary string decoding, the object variables are assumed to have a uniform search interval. One-point crossover and mutation operators are implemented in this study. In contrast, real-coded GAs work with variables directly, as they encode the solutions directly to the search space. Simulated binary crossover and polynomial mutation (Deb and Agrawal, 1995; Deb and Goyal, 1996; Deb, 2000) operators are used to create a child solution. A tournament selection (Deb, 2000) based on both constraint violation and objective function value is used for selecting individuals for crossover operation. The genetic operators are designed in such a way that they produce better solutions after generations. Sometimes, crossover and mutation may produce inferior solutions, but these solutions would disappear after passing through successive generations. To increase the efciency of the GA, it is necessary to preserve some of the better solutions for the next generation, without subjecting these to crossover and mutation operation. This is known as elitism. In this process, 5 to 10% of the best-performing members of the total population are preserved for the next generation, without applying crossover and a mutation operator. Constraint handling is a sensitive aspect of any constrained optimization algorithm. The constraint-handling capacity of GAs is better than classical optimization techniques because of their population-based approach (Deb, 2001). In most applications of constrained GA problems, the penalty approach is used to reformulate the problem into an unconstrained problem. Penalty parameters are needed to elevate the constraint violation values to the same order of the objective function value. It has been observed that, sometimes, the sensitivity of the solution is dependent on the value of the penalty parameter.
Hydrol. Process. 22, 842 853 (2008) DOI: 10.1002/hyp

C It

Rt

Et

8t 2 T

where St 1 is the storage at the end of the management period t 1, It is the inow into the reservoir during the time period t, Rt is the release from the reservoir during time period t, and Et is the evaporation loss during the time period t D 000006St . Simulation constraint. The functional relationship between the release quantities and the resulting concentrations as a function of time and space are specied as a binding equality constraint, represented by the numerical water quality simulation model: C R D f 9

where is the concentration vector and is the release C R vector. Water quality constraints. The maximum concentration of the lth constituent at the ith location at the end of the tth management period Cl is less than or equal i,t to the maximum permissible concentration at the ith location Cl . Cl is taken as the maximum value of i,max i,t concentration occurring over the whole depth at the ith location at the end of the tth management period: Cl Cl i,t i,max 8 l, i, t 2 T 10

Reservoir storage constraints. The ending storage at the end of each management period must lie between the minimum and maximum permissible storage values. The storage volume at the beginning of the rst management period S0 is specied: St,min St St,max 8t 2 T 11

where St,min is the minimum permissible storage at the end of time period t and St,max is the maximum permissible storage at the end of time period t. Reservoir release constraints. The specied release from the reservoir must be less than the maximum permissible release, which may be determined by the channel carrying capacity or by other factors. The minimum permissible release may be specied based on hydraulic or other considerations: Rt,min Rt Rt,max 8t 2 T 12

where Rt,min is the minimum permissible release at the end of time period t and Rt,max is the maximum permissible release at the end of time period t.
Copyright 2007 John Wiley & Sons, Ltd.

846

A. DHAR AND B. DATTA

Figure 1. Schematic representation of the developed methodology

Deb (2000) proposed a tournament-selection-based constraint-handling approach that can overcome the sensitivity of solutions to the penalty parameters. It utilizes the following criteria for tournament selection (Deb, 2000): (a) any feasible solution is preferred to any infeasible solution; (b) among two feasible solutions, the one having the better objective function value is preferred; (c) among two infeasible solutions, the one having the smaller constraint violation is preferred. The idea of comparing infeasible solutions in terms of constraint violation has implications from the optimization point of view. Preference for solutions with smaller constraint violation ensures convergence towards feasible solutions. This tournament selection method is utilized in this study to solve the constrained reservoir operation model. The elitist-GA-based optimization is initially utilized to generate the initial population of reservoir release values using a random generator. Each set of generated release values for each of the time periods of operation of the reservoir is sent to the simulation model, which has been implemented for the study area. The simulation model is then solved to obtain the resulting pollutant concentrations at each specied control point in the
Copyright 2007 John Wiley & Sons, Ltd.

river reach, at each time period of operation. These concentration values are utilized by the GA-based model to compute the corresponding tness function values for each member of the population. The tness function is a function of the specied objective function and constraint violation. In the next generation, a new population is created by crossover and mutation, from the members of the population in the earlier generation. In elitist-GA, the better performing individuals in terms of tness are preserved in the next generation. This procedure is repeated until the termination criterion is satised. The criterion used in this study is that the best solution reached has not improved over a prespecied number of generations. The resulting solution provides the optimal reservoir operation strategy over the time horizon considered. ILLUSTRATIVE APPLICATION AND PERFORMANCE EVALUATION The performance of the methodology developed is evaluated by applying the model to an illustrative study area. The hypothetical study area is a 55 km long river
Hydrol. Process. 22, 842853 (2008) DOI: 10.1002/hyp

OPTIMAL RESERVOIR OPERATION FOR DOWNSTREAM WATER QUALITY CONTROL

847

reach that is a part of the Middle Willamette River, Oregon, USA. A hypothetical reservoir is considered at the upstream end of the river reach. A single tributary drains into the river reach. The data needed to support the simulation model consists of three components: the river channel bathymetry, the meteorological conditions, and the boundary conditions, inows and temperatures. These hypothetical data were created using some of the data available in Willamette (2004). To evaluate the performance of the methodology developed, a management period of 10 days (maximum) starting from Julian day 158 to 168 is considered. Various management scenarios are considered for the evaluation purpose. In most of the scenarios the operation policy is obtained for a time horizon of 7 days, with management time steps of 1 day. In one of the scenarios, the operation policy is obtained for a total duration of 10 days. Figure 2 shows the top view, side view and crosssectional view of the reach, downstream to the reservoir. The tributary merges with the main river at segment 8. At the upstream boundary, a ow boundary condition is specied in which reservoir release is the inow to the reach. At the downstream boundary, an external head boundary condition is specied. The following assumptions are made: (a) the release from the reservoir is at a uniform rate over each management period; (b) tributary discharge is assumed uniform over the depth of the river at the location of junction; (c) short-term

reservoir operation policy is based on deterministic inow values with no forecasting errors incorporated. Four control points (10, 11, 12, 13) are chosen downstream of the tributary junction. The control points are at a sufcient distance upstream of the downstream boundary to avoid undue inuence of the assumed boundary condition on the concentration prole close to the boundary. In this illustrative problem, it is assumed that the tributary is discharging the pollutant nitrate plus nitrite (as N) at a high rate (Table I) throughout all management periods. Therefore, only one pollutant (l D 1) is considered in this example problem. The initial concentration of the constituent in the river reach and reservoir release is 014 mg l 1 . As per EPA standards (EPA, 2004), the maximum allowable nitrate plus nitrite concentration for drinking water is 10 mg l 1 . This concentration is taken as the basis for specifying the permissible concentration limit for the optimal reservoir release strategy. Bed elevation and initial water surface elevation (average) with respect to some external benchmark were specied as 178 m and 335 m respectively. In the downstream boundary, the nitrate plus nitrite concentration increased with time from 014 mg l 1 to 5 mg l 1 . The initial temperature is specied as 16 C in the river reach. Temperature is assumed to remain the same. The ULTIMATE scheme (Cole and Wells, 2003) is used for numerical simulation. Table II shows the parameter values used

Figure 2. (a) Top view of the downstream river reach; (b) side view of the river reach; (c) representative cross-sectional view of segment 3 Copyright 2007 John Wiley & Sons, Ltd.

Hydrol. Process. 22, 842 853 (2008) DOI: 10.1002/hyp

848

A. DHAR AND B. DATTA

Table I. Nitrate plus nitrite (as N) concentration incoming from tributary Management period Concentration (mg l 1 ) 1 900 2 900 3 1000 4 1000 5 1200 6 1200 7 1100 8 1100 9 900 10 1000

in the water quality simulation model. Different parameter values used in binary- and real-coded GAs are shown in Table III. The reservoir-operation time step as specied is different from the computational time step of the numerical simulation model. Once the management time step for the reservoir is specied, the simulation model automatically determines the time step for numerical solution as a fraction of the management time step. In the example problems solved here for performance evaluation, the specied initial storage is between the specied permissible minimum and maximum storages. This can be expressed as function of s such that: S0 D s St,max St,min C St,min 13

inow rates If are obtained as If D f Iinow 14

where f is the storage factor f 2 [0, 1] . In these performance evaluations, St,min is equal to 2008 106 m3 , St,max is equal to 6694 106 m3 , Rt,min is equal to 432 106 m3 day 1 and Rt,max is equal to 3456 106 m3 day 1 . The storage target Starget is specied as storage corresponding to s D 08 (i.e. 57568 106 m3 ). The specied tributary inows are shown in Figure 3.

where S0 is the initial storage and s is the storage factor s 2 [0, 1] . In the rst stage of performance evaluation, the reservoir inows Iinow are specied as shown in Figure 3, for Julian days 158 to 168. The model solutions are also evaluated by specifying inows as a fraction of the previously specied inow rates. For evaluation purpose, these
Table II. Different parameter values input to the simulation model Parameter Longitudinal eddy viscosity Longitudinal eddy diffusivity Maximum value of vertical eddy viscosity Mannings n Nitrate decay rate Denitrication rate from sediment Lower temperature for nitrate decay Lower temperature for maximum nitrate decay Unit m2 s m2 s m2 s
1 1 1

DISCUSSION OF RESULTS The performance of the proposed methodology was evaluated for different scenarios. These scenarios were chosen to demonstrate the operating policy for various realistic initial storage conditions and inow conditions. These scenarios incorporated different initial storages, different permissible maximum concentrations Cl , i,max different inows to the reservoir, and different time horizons of management. These different scenarios and the solution results are described here. The optimal release values obtained as solution are shown in Table IV and the corresponding storage values are shown in Table V. Scenario I. The parameter values specied in scenario I 1 are s D 01, f D 10, Cl i,max D 10 mg l , T D 7 days. In scenario I, the initial storage specied is 10% of the difference between the maximum and minimum permissible storage, over and above the minimum permissible storage St,min . Table V shows that the inow is contributing to increase in storage until the end of the fourth management period. At the end of the fth and sixth management periods, the storage values are reduced. Again,

Value 100 100 100 002 005 000 500 2500

day 1 m day 1
C C

Table III. Specied GA parameter values Parameter Population size Selection scheme Crossover rate Mutation rate Precision String per variable Crossover distribution index Mutation distribution index Binary coded 21 Tournament 0809 00101 Two places after decimal 12 Real coded 21 Tournament 0809 00101 2 100
Figure 3. Time-series of inows input to the optimization model Copyright 2007 John Wiley & Sons, Ltd. Hydrol. Process. 22, 842853 (2008) DOI: 10.1002/hyp

OPTIMAL RESERVOIR OPERATION FOR DOWNSTREAM WATER QUALITY CONTROL

849

Table IV. Optimal release values for the different scenarios and management periods 17 Release value (106 m3 ) Scenario I II III IV V VI 1 2443 2785 2439 3014 2848 2423 2 2284 2564 2300 2855 2661 2283 3 2014 2552 2014 2600 2572 2030 4 1947 2473 1930 2226 2436 1930 5 2264 2936 2257 2314 2912 2264 6 2302 2974 2302 2327 2953 2302 7 2081 2674 2052 2224 2647 2064

the storage value increases at the end of the seventh management period. The reason is that the inow coming to the reservoir until the end of the fourth management period is more than the release required to meet the downstream water quality standard. Therefore, only a portion of the inow is released to meet the downstream water quality requirement. The remaining portion of the inow increases the reservoir storage. Figure 3 shows that inows to the reservoir during the fth and sixth management periods are signicantly less. Also, during this time, the pollutant concentration in tributary discharge (Table I) increases and tributary inow decreases (Figure 3), resulting in a greater requirement of release from storage to meet the downstream water quality requirement. The downstream water quality standard is met by releasing water from storage at a rate greater than the inow. In the last management period, the inow to the reservoir increases and the concentration of tributary discharge reduces. Therefore, depletion of storage does not occur. The optimization model solutions were also validated by simulating the resulting concentrations for the reservoir releases (Table IV) as specied by the optimization model solution. As shown in Figure 4a, the maximum of the simulated concentration at all control points for each management period is close to the permissible maximum limit of 10 mg l 1 . Therefore, the optimization solutions satisfy the operational constraints as specied. These results show that the methodology developed is capable of evolving an operation strategy as intuitively expected. Also, the release policy actually satises the water quality constraints. Scenario II. The parameter values specied in sce1 nario II are s D 01, f D 10, Cl i,max D 8 mg l , T D 7 days. The only difference between scenarios I and II is a lower value of specied permissible concentration limit. This implies more stringent downstream water quality requirement. In this scenario, the permissible concentration limit is 8 mg l 1 . Therefore, the required release from the reservoir is larger in this scenario than in scenario I. This also results in a smaller storage value at the end of each management period. These values are shown in Table V. The inow to the reservoir until the end of the third management period is sufcient in magnitude to maintain the downstream water quality standard. For
Copyright 2007 John Wiley & Sons, Ltd.

the rest of the management periods, a part of storage contributes towards maintaining downstream water quality. It is also seen from the Figure 4b that the simulated concentrations simulated by using optimization-model-specied release values agree with the optimization model solution results. Scenario III. The parameter values specied in scenario 1 III are s D 01, f D 08, Cl i,max D 10 mg l , T D 7 days. In this scenario, the specied inows to the reservoir during each management period are assumed to be only 80% of the inows in scenario I. Other conditions are identical to scenario I. The maximum permissible concentration is 10 mg l 1 . Therefore, compared with scenario I, the storage depletion would be more in this scenario. The inow contributes to increasing the storage only for the rst management period, compared with the rst three management periods in scenario II. This is evident from Table V. Also, as expected, the ending storage at the end of the time horizon of operation is smaller than that in scenario I. Therefore, deviation from target storage is larger. The results shown in Figure 4c demonstrate the acceptability of the optimization model solutions. Scenario IV. The parameter values specied in sce1 nario IV are s D 08, f D 10, Cl i,max D 10 mg l , T D 7 days. In scenario IV, the specied initial storage is the target storage. This initial storage is much larger than that specied in scenario I. Therefore, because the inows specied for the majority of the management periods would be sufcient to satisfy the downstream water quality requirement without depleting storage, the optimal release policy is based mainly on the objective of maintaining the target storage value. In case of fth and sixth management periods, the storage value is reduced due to low inow into the reservoir and due to increased pollutant concentration in tributary discharge. The optimal release values (Table IV) are larger than those of scenario I for all management periods except the fth and sixth. These release values are the same as in scenario I, to maintain the downstream water quality requirement only. The simulated concentration values obtained using the optimal operation policies (Figure 4d) show that the prescribed releases are based on water quality requirements only for the fth and sixth management periods. In other management periods, the objective of meeting target storage dictates the releases, resulting in downstream concentration values much smaller than the specied permissible limit. Therefore, these solution results are again as rationally expected. Scenario V. The parameter values specied in sce1 nario V are s D 08, f D 10, Cl i,max D 8 mg l , T D 7 days. In scenario V, as in scenario IV, the initial storage is the target storage. However, the water quality constraint is more stringent. The permissible concentration is 8 mg l 1 and the inows specied are the same as in scenario I. The optimal operation policy obtained as a solution ensures that water quality requirements are satised and that a portion of inows in the rst three management periods contributes to the storage (Table V). The
Hydrol. Process. 22, 842 853 (2008) DOI: 10.1002/hyp

850

A. DHAR AND B. DATTA

Figure 4. Resulting concentrations at different segments: (a) scenario I; (b) scenario II; (c) scenario III; (d) scenario IV; (e) scenario V; (f) scenario VI

augmented storage can be utilized for increasing release rates in the fourth to seventh management periods, while minimizing the storage deviation from the target storage in these management periods. The imposition of a more stringent water quality standard (8 mg l 1 ) necessitates larger release rates in the fourth to seventh management periods, compared with that in scenario IV. The simulated downstream pollutant concentrations, as shown in Figure 4e, validate the optimal solution results. Scenario VI. The parameter values specied in scenario 1 VI are s D 08, f D 08, Cl i,max D 10 mg l , T D 7 days. In scenario VI, the specied initial storage is the target storage. The permissible concentration is 10 mg l 1 . The only difference between scenarios IV and VI is that the inows to the reservoir are only 80% of the inows in scenario IV. In this scenario, the storage is depleted to meet the water quality standard, although the release values (Table IV) are almost same as in scenarios I and III. In this scenario, satisfying the water quality constraint gets priority over meeting storage targets. This is same as in scenarios IV and V. As expected, the ending storage at the end of the time horizon of operation is smaller than that in scenario IV. Figure 4f shows the pollutant concentrations simulated by using the specied

optimal releases obtained as solutions of the optimization model. The simulated concentrations satisfy the imposed water quality constraints. Scenario VII. The parameter values specied in sce1 nario VII are s D 08, f D 10, Cl i,max D 10 mg l , T D 10 days. In scenario VII, the time horizon for solution of operation model is 10 days instead of 7 days. Other conditions are same as in scenario IV. As shown in Figure 5, operation policies obtained as a solution are not identical for scenarios IV and VII. This is expected, as actual optimality of any policy model depends on the time horizon considered. Shorter time horizons can result in solutions that are myopic in nature, and not optimal for larger time horizons. Therefore, deviation from storage target at the end of the seventh management period is larger in scenario VII than that in scenario IV. The actual optimality of the optimal solutions obtained by solving the optimization problem is an important issue. The search technique used by the optimization algorithm may or may not be able to reach the actual or global optimum. The real-coded GA was used for obtaining the solutions reported here. In order to test the efciency of this algorithm, binary-coded GA solutions were also obtained. All the results presented

Table V. Optimal storage values for the different scenarios and management periods 17 Storage value (106 m3 ) Scenario I II III IV V VI S0 106 m3 247660 247660 247660 575680 575680 575680 1 248280 247938 247668 575700 575875 575684 2 248828 248206 247622 575667 576027 575645 3 249371 248211 247650 575606 575993 575639 4 249813 248127 247628 575747 575926 575596 5 249800 247440 247167 575664 575245 575110 6 249773 246741 246683 575593 574548 574606 7 250021 246398 246491 575678 574211 574381

Copyright 2007 John Wiley & Sons, Ltd.

Hydrol. Process. 22, 842853 (2008) DOI: 10.1002/hyp

OPTIMAL RESERVOIR OPERATION FOR DOWNSTREAM WATER QUALITY CONTROL

851

Figure 5. Release values for scenarios IV and VII

thus far are based on real-coded GA solutions. The solutions obtained by using a binary-coded GA were inferior to those for real-coded GA for comparable population sizes and number of generations. A typical comparison is shown in Figure 6. It is evident that the real-coded GA prescribes a more efcient operation policy. This can be veried by the fact that some of the prescribed releases obtained by using the binary-coded GA are larger than those specied as a solution of the real-coded GA. However, the releases prescribed by the real-coded GA, although smaller, meet the water quality constraints. The objective function value for the binary-coded GA is also larger. However, it cannot be ruled out that, with very large population size and very large number of generations, a global optimum solution may be reached even with a binary-coded GA. The GA adopted is the elitist GA with tournament selection (Deb, 2000) for handling constraints. No doubt, the convergence of the solution resulting in an optimal solution depends on various factors, including the assigned parameter values. The population size, the total

number of generations, and/or the stopping criteria may determine the optimality or non-optimality of the solution results. However, these numbers are problem specic and difcult to generalize. Figure 7a and b shows that the best objective function value is increasing for rst few generations. This is because the initial solutions are generally infeasible, which may result in very low objective function value in a minimization problem. In subsequent generations, the best objective function value may actually deteriorate, even with an elitist search. This is because constraint violations are expected to decrease as the best solution marches towards feasible solutions in the subsequent generations. Figure 7a and b shows the best objective function value against the number of generations. The peaks correspond to the rst identied feasible solution, which, no doubt, may represent a sudden increase in the objective function value for a feasible solution. The objective function values preceding these peaks do not correspond to a lower objective function value (in minimization) because of the nature of constraint handling scheme adopted here. This scheme ensures that the better of two solutions from a pair of solutions chosen by tournament selection is based on lower constraint violation, if all individuals are infeasible. The stopping criterion for terminating the GA in this study is based on convergence. This stopping criterion ensures that the solutions are feasible and that the numbers of generations are sufcient so that the best solution in terms of tness do not improve over a large number of generations. It is evident (Figure 7) that the numbers of generations were large enough to meet the stopping criterion. Most of the solutions required 40 h of CPU time (24 GHz processor, 256 MB RAM) to complete 500 generations with population size of 21 and with one elitist member. This time requirement depends upon the

Figure 6. Comparison of optimal releases obtained using real- and binary-coded GAs for scenario I Copyright 2007 John Wiley & Sons, Ltd.

Figure 7. Variation of best objective function value with number of generation for (a) scenarios I, II, III and (b) scenarios IV, V, VI Hydrol. Process. 22, 842 853 (2008) DOI: 10.1002/hyp

852

A. DHAR AND B. DATTA

processor speed and RAM. This CPU time requirement is also dictated by the number of generations completed. It is to be noted that a large portion of the CPU time is necessary to solve repeatedly the chosen numerical simulation model, CE-QUAL-W2. The optimization model requires these solution results (i.e. resulting pollutant concentrations) as inputs from the numerical model. The external linking of the pollutant transport numerical simulation model may increase the CPU time requirement. However, this makes it possible to incorporate various complex transport processes and sophisticated numerical techniques to obtain meaningful optimal operation policies without increasing the complexity of the optimization model. In the performance evaluation results presented here, nitrate plus nitrite (as N) is specied as the pollutant. This choice is purely for illustrative purposes. Any other suitable pollutant may have been chosen for illustration. The methodology developed is not dependent on the choice of the pollutant, so long as the pollutant transport processes can be simulated by the externally linked simulation model. It is even possible to include pollution criteria for multiple pollutants, using the generalized formulation of the optimization model as proposed here. It is also possible to extend the optimal operation model to a system of multiple reservoirs. The short-term operation policy is based on the objective of minimizing the deviation from a specied storage target. The storage targets should be determined based on various economic considerations. In real-life cases, the specication of a storage target may require extensive simulation and analysis. The results presented here are purely for illustrative purposes to establish that the solution results are acceptable. These results do not in any way establish the range of validity of the proposed methodology. Also, these evaluations for demonstrating potential applicability are limited in scope. The input variable values were chosen such that the variation in solution results for different scenarios can be highlighted. Also, the specied permissible concentration and the concentration of the pollutant reaching the main river reach from the tributary are only for illustrative purposes.

different management scenarios for a specied hypothetical system. The performance of the methodology developed is evaluated for different permissible water quality standards, different initial storage conditions, different inow rates, and different time horizons. The solution results obtained are also validated by solving the water quality simulation model using the prescribed optimal release rates. Although the performance evaluations are very limited in scope, the potential feasibility of using the methodology developed to solve complex non-linear reservoir system operation models is demonstrated. Further rigorous evaluations are no doubt necessary to establish the general applicability of the methodology developed. LIMITATION AND FUTURE SCOPE The proposed methodology inherits the limitation of the simulation model itself, although it is very easy to replace the simulation model by any other suitable simulation model. The methodology developed did not incorporate the stochastic nature of inows or the inow forecasting errors. Incorporation of uncertainties and forecast errors would be an interesting extension. In the example problem, seepage or groundwater ow has not been included. It is true that groundwater seepage is an important component of the water balance. The study considers the operation of a single river reservoir system with a single objective. The methodology developed can be extended to a system of reservoirs for more efcient synergistic operation. No doubt, the increase in number of reservoirs with longer time horizon of management will increase the burden of computation. The CPU time required even for a small problem is quite large. Development of parallel code or use of metamodels (e.g. ANNs) may be very useful in reducing the CPU time. This would make it feasible to solve larger and more complex real-life optimal reservoir system operation problems. Incorporation of multiple objectives of operation would be a useful extension.

REFERENCES Cai X, McKinney DC, Lasdon LS. 2001. Solving nonlinear water management models using a combined genetic algorithm and linear programming approach. Advances in Water Resources 24: 667 676. Carmichael JJ, Strzepek KM. 2000. A multiple-organic-pollutant simulation/optimization model of industrial and municipal wastewater loading to a riverine environment. Water Resources Research 36: 1325 1332. Celeste AB, Suzuki K, Kadota A. 2004. Genetic algorithms for realtime operation of multipurpose water resources systems. Journal of Hydroinformatics 6: 1938. Chang FJ, Chen L. 1998. Real-coded genetic algorithm for rule-based ood control reservoir management. Water Resources Management 12: 185198. Chaves P, Kojiri T. 2003. Multi-objective storage reservoir operation under uncertainty. Disaster Prevention Research Institute Annuals, Kyoto University 46B: 919 928. Chaves P, Kojiri T, Yamashiki Y. 2003. Optimization of storage reservoir considering water quantity and quality. Hydrological Processes 17: 2769 2793. DOI: 101002/hyp.1433. Chen L. 2003. Real coded genetic algorithm optimization of long term reservoir operation. Journal of American Water Resources Association 39: 1157 1165. Hydrol. Process. 22, 842853 (2008) DOI: 10.1002/hyp

CONCLUSIONS A methodology is developed for optimal short-term operation of reservoir systems to control downstream water quality. The operation policy is obtained as the solution of a linked simulation optimization model. It links an optimization algorithm (i.e. an elitist GA) and a surface water quality simulation model (namely CE-QUAL-W2) for reservoir operation with water quality constraints. The linked simulation optimization methodology considers a single objective of operation, i.e. minimization of deviation from a target storage. It incorporates the non-linear relationship between reservoir release and downstream water quality. The methodology developed is solved for
Copyright 2007 John Wiley & Sons, Ltd.

OPTIMAL RESERVOIR OPERATION FOR DOWNSTREAM WATER QUALITY CONTROL Cole TM, Wells SA. 2003. CE-QUAL-W2: a two-dimensional, laterally averaged, hydrodynamic and water quality model, version 3 1. Instruction Report EL-03-1, US Army Engineering and Research Development Center, Vicksburg, MS. http://www.cee.pdx.edu/w2 [April, 2004]. Dai T, Labadie JW. 2001. River basin network model for integrated water quantity/quality management. Journal of Water Resources Planning and Management 127: 295305. Dandy G, Crawley P. 1992. Optimum operation of a multiple reservoir system including salinity effects. Water Resources Research 28: 979 990. Datta B, Soraganvi VS. 1996. Water quality control and hydropower generation as conicting objectives of optimal reservoir operation. In Proceedings of International Conference on Aspects of Conicts in Reservoir Development and Management, City University, London. Deb K. 2000. An efcient constraint handling method for genetic algorithms. Computer Methods in Applied Mechanics and Engineering 186: 311338. Deb K. 2001. Multi-Objective Optimization using Evolutionary Algorithms. Wiley. Deb K, Agrawal RB. 1995. Simulated binary crossover for continuous search space. Complex Systems 9: 115 148. Deb K, Goyal M. 1996. A combined genetic adaptive search (GeneAS) for engineering design. Computer Science and Informatics 26(4): 30 45. Donigian Jr AS, Imhoff JC, Ambrose Jr RB. 1995. Modeling watershed water quality. In Environmental Hydrology, Singh VP (ed.). Kluwer. EPA. 2004. 2004 Edition of Drinking Water Standards and Health Advisories. EPA-822-R-0 4005. Esat V, Hall MJ. 1994. Water resources system optimisation using genetic algorithms. In Hydroinformatics 94, Proceedings of 1st International Conference on Hydroinformatics. A. A. Balkema: Rotterdam; 225 231. Fahmy HS, King JP, Wentzel MW, Seton JA. 1994. Economic optimization of river management using genetic algorithm. In ASAE 1994 International Summer Meeting, American Society of Agricultural Engineers, St Joseph, MI; paper no. 943034. Goldberg DE. 1989. Genetic Algorithm in Search, Optimization, and Machine Learning, Addison-Wesley. Gunduz O, Soyupak S, Yurteri C. 1998. Development of water quality management strategies for the proposed Isikili reservoir. Water Science and Technology 37: 369376. Hayes DF, Labadie JW, Sanders TG, Brown JK. 1998. Enhancing water quality in hydropower operations. Water Resources Research 34: 471 483. Hejzlar J, Sustrova H, Mikesova P, Ruzicka M, Kafkova DA. 2002. Scenario study for reservoir water quality management. In ResLim 2002, Extended Abstracts, 4th International Conference on Reservoir Limnology and Water Quality, Cesk Bud jovice, 1216 August, e e Institute of Hydrobiology Academy of Sciences of the Czech Republic, Faculty of Biological Sciences University of South Bohemia; 121124. Holland JH. 1975. Adaptation in Natural and Articial Systems. University of Michigan Press. Huang WC, Yuan LC, Lee CM. 2002. Linking genetic algorithm with stochastic dynamic programming to the long-term operation of multireservoir system. Water Resources Research 38: 1304. DOI: 101029/2001WR001122. Ikebuchi S, Takasao T, Kojiri T. 1982. Real-time operation of reservoir systems including ood, low ow and turbidity controls. In Experience in Operation of Hydrosystems, Unny TE, McBean EA (eds). Water Resources Publications: Littleton, CO; 25 46.

853

Jaworski NA, Weber Jr WJ, Deininger RA. 1970. Optimal reservoir releases for water quality control. Journal of the Sanitary Engineering Division, Proceedings of American Society of Civil Engineers 96(SA3): 727 742. Ko IH. 1997. Integrated river basin operational planning considering water quantity and water quality. PhD dissertation, Colorado State University, Fort Collins, CO. Kojiri T. 1987. Systems analysis in multi-dam reservoir systems including water quantity and quality controls. In Environmental Geotechnics and Problematic Soils and Rocks, Balasubramanaiam AS, Chandra S, Dergado DT, Nutalaya P (eds). A. A. Balkema: Rotterdam; 25 37. Labadie JW. 1997. Reservoir system optimization models. Water Resources Update, The Universities Council on Water Resources (UCOWR) 108: 83110. Labadie JW. 2004. Optimal operation of multireservoir systems: state-ofthe-art review. Journal of Water Resources Planning and Management 130: 93111. Loucks DP, Jacoby HD. 1972. Flow regulation for water quality management. In Models for Managing Regional Water Quality, Dorfman R, Jacoby HD, Thomas HA (eds). Harvard University Press: Cambridge, MA; 362 431. Nandalal KDW, Bogardi JJ. 1995. Optimal operation of a reservoir for quality control using inows and outows. Water Science and Technology 31: 273 280. Ndiritu JG. 2003. Reservoir system optimization using a penalty approach and a multi-population genetic algorithm. Water SA 29: 273280. Nicklow JW, Ozkurt O, Bringer Jr JA. 2003. Control of channel bed morphology in large-scale river network using genetic algorithm. Water Resources Management 17: 113132. Olivera R, Loucks DP. 1997. Operating rules for multireservoir systems. Water Resources Research 33: 839 852. Orlob GT, Simonovic S. 1982. Reservoir operation for water quality control. In Experience in Operation of Hydrosystems, Unny TE, McBean EA (eds). Water Resources Publications: Littleton, CO; 263 285. Rajbhandari HL. 1995. Dynamic simulation of water quality in surface water systems utilizing a lagrangian reference frame. PhD dissertation, University of California, Davis. Sharif M, Wardlaw R. 2000. Multireservoir systems optimization using genetic algorithms: case study. Journal of Computing in Civil Engineering 14: 255 263. Simonovic S, Orlob GT. 1982. Optimization of reservoir operation for water quality management. In Environmental Systems Analysis Management, Rinaldi S (ed.), North Holland, Amsterdam, 311 327. Simonovic S, Orlob GT. 1984. Risk-reliability programming for optimal water quality control. Water Resources Research 20: 639646. Wardlaw R, Sharif M. 1999. Evaluation of genetic algorithms for optimal reservoir system operation. Journal of Water Resources Planning and Management 125: 2533. Willamette . 2004. http://www.cee.pdx.edu/w2/Willamette [April 2004]. Willey RG, Smith DJ, Duke Jr JH. 1996. Modeling water-resources systems for water quality management. Journal of Water Resources Planning and Management 122: 171179. Wu R-S, Sue W-R, Chen C-H, Liaw S-L. 1996. Simulation model for investigating effects of reservoir operation on water quality. Environmental Software 11: 143150. Wurbs R. 1993. Reservoir-system simulation and optimization models. Journal of Water Resources Planning and Management 119: 455472. Yeh WW-G. 1985. Reservoir management and operation models: a stateof-the-art review. Water Resources Research 21: 1797 1818.

Copyright 2007 John Wiley & Sons, Ltd.

Hydrol. Process. 22, 842 853 (2008) DOI: 10.1002/hyp

You might also like