A DSS for planning and managing water reservoir systems

Environmental Modelling & Software 18 (2003) 395–404www.elsevier.com/locate/envsoft

A DSS for planning and managing water reservoir systems

R. Soncini-Sessa∗, A. Castelletti, E. WeberDipartimento di Elettronica e Informazione, Politecnico di Milano, Piazza L. da Vinci, 32, I-20133 Milan, Italy

Received 28 July 2002; received in revised form 4 December 2002; accepted 13 February 2003

Abstract

The technologies and methods for integrated planning and management of water resource systems have matured considerablyover the past decades. However, relatively few of them have been actually and regularly applied in real world decisional processes.We feel this is essentially due to a general lack of engagement of stakeholders and decision makers at every stage of the decisionalprocess. Innovative methodologies and tools to improve participation are presented, with focus on water reservoir systems. 2003 Elsevier Science Ltd. All rights reserved.

Keywords:Water management and planning; DSS; Negotiation; Management policies

1. Introduction

It is very probable that the century we have justentered will be remembered as the “century of water”,since water will be by far the scarcest resource, the avail-ability of which will constrain the economy of nations.In the last 50 years water demand in the world hastrebled and is still increasing sharply every year as aresult of population growth, increase in householdincome, and irrigation development. Approximately 70%of surface water and groundwater is claimed by irrigatedagriculture, to produce 40% of worldwide food needs(Brown, 2001). Even though demand for water isincreasing, rare are the countries in which its availabilityremains constant; more frequently, the exploitation iswell beyond the renewal rate of the resource. Degradingwater quality further reduces the availability of fresh-water suitable for domestic and agriculture use andincreases the cost of treatment and reuse of water.Driven by these challenges, the last years have seen awide and growing interest towards the development oftools and techniques for integrated planning and man-agement of these systems (see, e.g.SFWMG, 1987;Simonovic and Savic, 1989; Loucks, 1990). While in the1960s, the planning was based on the assumption that

∗ Corresponding author.E-mail address:[email protected] (R. Soncini-Sessa).

1364-8152/03/$ - see front matter 2003 Elsevier Science Ltd. All rights reserved.doi:10.1016/S1364-8152(03)00035-5

water was an infinite resource and the main concern wasits allocation and distribution, now the approach mustbe oriented to sustainability and must be more holistic.

Nonetheless, technology alone is not enough, if theprojects that it suggests actually remain unrealized: infact the accomplishment is quite often politicallythwarted by the disagreement of all the stakeholders thatare excluded by the benefits or that fear damage from theproposed actions. To overcome this risk, it is essential toassess in detail the likely impacts of a project not onlyon those objectives for which it has been conceived anddesigned, but also on all the sectors that it may influence.The human role in the decisional process must be sup-ported and valued by exploiting the power of models todetermine the effects of alternative projects with lowcost and low impact computer-based experiments. Theselected course of action should emerge out of a debate,where all the stakeholders are involved. To reach thisgoal a formalized planning procedure was required, aprocedure able to determine a solution that is equitable,efficient and sustainable: the Environmental ImpactAssessment (EIA) procedure emerged as the most appro-priate. Indeed in the last decades many countries haveenacted legal frameworks for EIA, by specifying rules,methodologies and guidelines for its application in dif-ferent planning situations.

The water resource systems we will consider in thispaper are storage systems, including multipurpose reser-voirs or regulated natural lakes, catchments, channels,

396 R. Soncini-Sessa et al. / Environmental Modelling & Software 18 (2003) 395–404

rivers, streams, water users, etc., all of them intercon-nected by the downstream flow of water. Water supplyfor agriculture and power generation are usually con-sidered the two main purposes of such systems; however,the water collected by these facilities may serve severalother scopes, such as navigation, recreation, flood pro-tection, downstream river quality conservation, etc.Therefore, any decision taken over these systems directlyor indirectly affects a wide range of stakeholders.

2. EIA procedure

The EIA procedure can be formalized in a five-stageprocess as outlined in Fig. 1. The scheme draws a con-ceptual map of phases to be followed by the system ana-lyst (SA). More in detail, given a structural configurationof the system (existing or proposed) and on the basis ofthe strategic goals pursued, the SA should elaborate andcompare alternative planning proposals with the follow-ing procedure:

1. Indicator specification: the strategic goals are trans-lated into operational criteria and the latter into physi-cal and economical indicators, which can be thenquantitatively computed. For example, the indicatorsmay evaluate the performances in supplying water tocivil, agricultural and industrial users, the complianceto river quality standards, and flood control targets.Since the operational criteria reflect concerns and pri-orities of the stakeholders, they should be defined byinteracting with them.

2. Model identification: the components (catchments,reservoirs, channels, water users, etc.) associated witha given structural configuration of the water systemare described by mathematical models. The choice ofthese models as well as the degree of detail dependson both the selected indicators and the alternatives tobe evaluated, thus there exists a recursion among thisphase and phases 1 and 3.

Fig. 1. The EIA procedure scheme.

3. Alternative identification: all the feasible structuraland/or normative actions are first quantified and thencombined in all possible way. Each combination con-stitutes a planning alternative.

4. Alternative evaluation: for each alternative, the valuesof the indicators are assessed by simulating thebehavior of the system.

5. Alternative comparison and negotiation: the prefer-ence systems of the stakeholders are identified. Then,the alternative proposals are compared according tomulti-attribute and negotiation aid techniques.Finally, a sensitivity analysis is performed to checkthe robustness of the ranking with respect to theuncertain and subjective elements. All these actionshelp the decision maker (DM) to select a compro-mise decision.

The practical application of the above procedure posestwo fundamental concerns: first, the SA should widelyinteract with all stakeholders during the phases of indi-cator specification, model identification and alternativeevaluation; second, the information used at each stageshould be obtainable from the previous stages and madeavailable for the successive. More in detail this infor-mation must be complete, shared, transparent, easilyobtainable, well structured and flexible. The problem isfurther complicated by the non-strictly serial successionof stages and by the diffuse presence of recursions.Hence, the effective implementation of the abovescheme requires the support of a computer-based system,a Decision Support System, which provides the tools tointegrate all the phases in a unique framework, facilitatesparticipation, and ensures that the information has theabove mentioned properties.

The proposed EIA scheme is a procedure of generalvalidity. It has been originally introduced for a delimitedclass of problems, but it may be applied to a broad rangeof planning contexts. Its use for water system planningposes several challenging and significant issues. In thispaper, we will analyze four of them and propose for eachone some innovative approach. We will conclude thepaper by presenting the architecture of a DSS which sat-isfies the above mentioned requirements and imbeds theinnovative proposals.

3. Dynamic and control

EIA analysis as well as the Multi-Attribute ValueTheory (MAVT) (Keeney and Raiffa, 1976), upon whichit is based, have been traditionally developed and areusually adopted to evaluate projects on static systems(or systems that are assumed to be static for modellingreasons). Then they implicitly consider decisions asplanning actions, that is una tantumdecisions, that per-manently modify the system configuration. This way of

397R. Soncini-Sessa et al. / Environmental Modelling & Software 18 (2003) 395–404

doing does not lend itself to the intrinsically dynamicnature of water reservoir systems. To better clarify thisissue let us consider, e.g. the Verbano water system,located on the Italian–Swiss border (Soncini-Sessa et al.,2000). Its storage is primarily used to provide water sup-ply to two wide irrigation districts and several hydro-power plants, but it also influences a plethora of othersocio-economical and environmental factors (lacustrineand riverine water quality, upstream and downstreamflood control, lacustrine navigation and tourism, and soon). Obviously, the volume of water stored in the lake(i.e. the state of the system) is time-variant. At each timeinstant (i.e. every morning) the DM must decide howmuch water has to be released in the next 24 h. Clearly,the volume of water that can be released depends on thevolume of water currently available into the lake. Therelease decision influences the lake storage of the nextday and hence influences the next release decision.Decisions are thus concatenated after each other andrecursive, that is they are management actions. Theseactions can be formulated in a rational way by designinga management policy (see definition in the nextparagraph), that can conceptually be included in the EIAat the same level of the other planning actions. Straight-forwardly, a planning alternative is constituted by a tri-ple: structural actions (i.e. dredging of the lake outlet),normative actions (i.e. review of the regulation range andimposition of a minimum instream flow) and one man-agement policy. The identification of the alternative inphase 3 should therefore include the “quantification” , i.e.the design, of the management policies too.

4. The management policy

It is time to give a more formal insights on manage-ment policies. We will make reference to the Verbanocase to be less abstract. Known the current lake storagest at time step t, the release decision ut is given by:

ut � mt(st) (1)

where m.(·) is a succession, generally periodic of periodT, of monotone non-decreasing functions of st, namedcontrol laws. A management policy p is a succession[m0(·),m1(·),…,mT�1(·),…] of control laws. Given a cou-ple of structural and normative actions, a managementpolicy is synthesized by solving an optimal control prob-lem, defined as:

minp

limt→�

Crite1,…,et+1��n

i � 0

liZi(p)� (2a)

Zi(p) � �t

t � 0

git(st,ut,et+1) (2b)

st+1 � ft(st,ut,et+1) (2c)

ut�Ut(st) (2d)

et+1 � ft(et+1�st,ut) (2e)

ut � mt(st) (2f)

p � [m0(·),m1(·),…,mT�1(·)] (2g)

where all the functions are periods T. The weights li

(i = 1,…,n) expresses the relative importance of n indi-cators Zi(p), chosen among the ones identified in phase1, each one of which is the sum over time of step costsgi

t(·). The set Ut(st) is the set of feasible controls (thatmay vary with the storage st); ft(et + 1�st,ut) is the con-ditional probability distribution of the disturbance et + 1

(e.g. the inflow to the lake) and Crit is a suitable criterion(i.e. a statistic) to filter the disturbance (see Yakowitz,1982; Soncini-Sessa et al., 1991). The structural–norma-tive actions are embedded in the constraints (2c) and ofthe control problem. Thus, given a pair of structural–normative actions, several (theoretically infinite) differ-ent management policies can be devised, by varying theweights li. The set of these policies constitutes the Par-eto boundary of the problem, that is the set of non-domi-nated policies with respect to the management objec-tives.

When the water system is more complex than the Ver-bano water system (e.g. it has more reservoirs), its stateis a vector xt, that contains the states of all the dynamicalcomponents within the system. Then the managementpolicy can be synthesized by a problem analogous toproblem (2), where st is substituted by xt and ut is thevector of all the decisions that must be daily assumedin the system.

5. A two-level decision approach

The analysis of the decisional organization of a WaterAgency leads towards the idea of a multi-level DSS. Infact, in a Water Agency, one can usefully distinguishdifferent levels of decision making. Anthony (1965) con-siders three levels: strategic planning, management con-trol, and operational control (denoted in the followingsimply as planning, management and control) (see Fig.2). The planning level has to do with the strategic goals

Fig. 2. The two-level decisional process.


to be pursued in the system management and with theways to achieve them: management policies are designedat this level. The policies set up at the planning levelcan be expressed either by closed form rules, resultingfrom the solution of off-line optimal control problems,or by on-line optimal control schemes. The scope of themanagement level is the utilization of the resources in aneffective way in the short and medium term, accordingto the directive issued by the planning level and to theparticular situation the DM is facing (e.g. the out oforder of a power plant). This is the level were the man–machine interaction is more tight and where conflicts arelikely to arise. The control level has to implement themanagement decisions in the real time operation and itis generally highly structured, so that fixed rules can beapplied. Thus, the control level is served by a controller,while the other two levels need two separate DSSs:DSS/M and DSS/P where M stands for management andP for planning. The planning module DSS/P deals withthe choice of the alternatives and its outputs are planningdecisions, management policies, and models. All of themconstitute inputs of the management module DSS/M.The interaction among the SA, the DM, and third parties(stakeholders) takes place in the DSS/P at the planninglevel and in the DSS/M at the management level. Thescheme of Fig. 1 can be embedded in this two-levelstructure, thus obtaining the scheme in Fig. 3.

The two-level structure embodies an interesting skill:the planning tools are naturally and inexpensivelyupdated. In the usual planning approach, once the plan-ning action is over, the planning tools are abandoned;thus, when later on a new planning action is required,the construction of the planning tools must starts againfrom scratch, with a loss of money, time and coherence.On the contrary, in the two-level DSS the daily use ofthe DSS/M implies that the data base and the systemmodel are regularly updated, then, since the DSS/Pshares them with the DSS/M, the DSS/P is automati-cally updated.

Fig. 3. The phases of the two-level decisional process.

6. Set-valued policies

The definition of a management policy spawns fromControl Theory, which was originally developed to con-trol electro-mechanical systems. In such systems, thepolicy must allow for control in absence of human inter-vention, i.e. it must produce an automatic control.Examples of such policies are the autopilot of a jetlineror the speed controller of an engine. Since control mustbe automatic, at every time step the policy must returnone and only one control decision: therefore, the policymust be a point-valued function.

On the contrary, in the management of water systems,the policy does not operate directly on dams and weirs,but proposes a control decision to the DM. It will be upto the DM to accept the decision suggested. Note in factthat the optimality of a management policy does not referto the real world complexity, but to a simplification ofthe real world, that is to a scenario. The scenarioincludes all the approximations that are behind the usedmodels and the selected objectives. For instance, thestructure of the downstream water distribution networkis usually heavily simplified, the catchment model(which generates the reservoir inflows) is often modelledas a purely stochastic process. These simplifyingassumptions are justified, from the SA viewpoint, sincethey are needed to reduce the optimal control problemto a solvable formulation, given the current mathematicaland hardware tools. However, at the same time they areperceived as deceiving by the DM. Consequently, itwould be more effective if the policy, rather than propos-ing a single decision value ut, lists a set Mt (safe controlset) of decisions that are equivalent (i.e. which guaranteein the long run the same performance of the policy) fromthe point of view of management objectives and for themodels adopted to represent the real world (for detailssee Aufiero et al., 2001a, b). The use of a set-valuedpolicy would therefore allow the DM to choose at timet a decision among the optimal ones by taking intoaccount new facts: e.g. secondary objectives (notincluded in the objective Zi of problem (2a)) that appearto her/him as particularly relevant at time t or the currentsituation (e.g. heavy rainfall on the catchment, powerplant outages, etc.) that has not been considered in thescenario adopted in the policy design phase. In this way,the level of confidence of the DM in the policy is widelyimproved. A straightforward consequence of the adop-tion of set-valued policies is that the controller wherethe policy is implemented (see phase 6 in Fig. 3) mustbe a Decision Support System itself (that is the reasonbehind the acronym DSS/M), in order to enable the DMto evaluate and compare the effects of alternative choicesof ut in Mt.


7. Systems and performances

We have seen how a planning alternative can be ident-ified and how the design of set-valued management poli-cies plays a central role in this process. We come nowto the questions posed by the evaluation of each alterna-tive. The stakeholders with the same concerns and pri-orities are grouped in one sector, to which is associatedan index. This index is a value expressing the stake-holders’ satisfaction for a particular alternative. Theindexes are chosen in phase 1 and used in the successivephases to compare the alternatives. For a given sector,the evaluation of the corresponding index is carried outin cooperation by the SA and the so-called sector expert.This is usually a technician who is expected to supportthe stakeholders by translating their qualitative judg-ments into quantitative evaluations. Theoretically thesector expert could directly settle the index value foreach alternative on the basis of expertise (Fig. 4a). How-ever, this approach frequently leads to too subjective andhardly acceptable evaluations, that might further becomeimprecise when faced with a large number of alterna-tives. It would be then more appropriate to formalize aprocedure to compute the index values automatically(Fig. 4b). This procedure should reproduce the expertskills and be identifiable through interviews.

The dynamic nature of a water reservoir systemimplies that its evolution over a given time horizon maybe completely described by the trajectories of the systemoutputs (i.e. reservoir water storages and releasedecisions). Hence, an alternative modifies the trajectories

Fig. 4. More and more articulated ways of structuring an index.

of its outputs. The process of evaluating an index canthus be thought of as a two step process (Fig. 4c):

1. the simulation of the controlled system over a giventime horizon, by using the mathematical model ofthe system;

2. the computation of the index as a function of the tra-jectories computed in the first step.

However, the direct identification of an index throughthe simple observation of the system trajectories couldagain be a hard task for the sector expert. To furthersimplify the problem, one or more intermediate indi-cators may be conveniently added between the trajector-ies and the index (Fig. 4d). These indicators must meeta precise condition: each one must be expressed by aseparable function of the system variables, i.e. it mustbe the temporal aggregation of instantaneous indicators(the step costs gi

t(·) in Eq. (2b)), such that the one at timet only depends on values of variables at the same timet. Note that this condition implies that each step cost attime t cannot be dependent on the value of any step costat previous times, i.e. the step costs cannot be state vari-ables of any system. This condition seems to be parti-cularly restrictive, but it can always be met by suitablyenlarging the state of the system model (e.g. if the stepcost depends upon its value at a previous time, it cansimply be included among the state variables by addinga transition function that expresses its dynamics to themodel).

More in general, the indicators should be easilyobtainable given the trajectories and should make it eas-ier for the expert to relate them to an index. This latterstep is usually covered by the well known MAVTthrough the definition of utility functions and therefore itwill not be discussed further. We focus instead on modelconstruction and indicator definition.

Both these steps require a wide interaction betweenthe SA, the DM and the stakeholders, since they should“ feel” the models adopted as well as the indicators selec-ted. Only in this way does the decisional process leadto evaluations that will be perceived as “ trustable” : adirect involvement of the decision maker herself in themodelling process is the unique way to make crediblemodels: these can in fact be built only by people whoare familiar with both the problem and the institutionalsetting in which the problem is to be addressed (Louckset al., 1985).

8. Models and indicators

The components of a water reservoir system are usu-ally modelled by means of models, either simple “black-box” or complex physically based models, that can berepresented by the following equation:


ct+1 � ft(ct,ut,wt,et+1) (3a)

yt � ht(ct,ut,wt,et+1) (3b)

where ct is the state variable, yt is the output, wt is theexogenous input to the component (i.e. generally the out-put of other components) and et + 1 is a disturbance. Allthese models are markovian models. Once the state tran-sition function ft(·) and the output function ht(·) havebeen properly structurally identified, they can be easilycalibrated with the well known techniques of SystemIdentification Theory (Ljung, 1987). Consider forinstance the catchment: it can be described using a stoch-astic autoregressive model of order q, where ct is thestate of the catchment, yt represents the outflow (i.e. thereservoir inflow) and et + 1 is a disturbance that representsrainfall, temperature, etc. Both ft(·) and ht(·) are linearfunctions, whose parameters values might be easilyidentified from historical inflow series.

However, it may occur that the physical and socio-economical relations that underlie some of the compo-nents are poorly known and/or it is comparativelyexpensive to obtain raw data to characterize them better.In these cases, it might be difficult to identify the twofunctions ft(·) and ht(·). The SA is thus forced to formu-late unrealistic simplifications, the majority of which arebased on the assumption that the processes are wellknown and deterministic. But this assumption is oftentoo reductive. Consider again the Verbano water system.A suitable way to lower the water demands of the irri-gation districts is to reduce the inefficiencies of the irri-gation systems. A lower downstream water demandshould in fact reduce the water storage required to reacha given level of supply satisfaction and, as a conse-quence, induces a reduction of flood risk on the lakeshores. Thus, the DM might wonder whether and whatlevel of financial incentive would encourage the farmersto adopt more water-efficient irrigation systems. Intuit-ively, what the DM expects is that the farmers’ bent tomodify their irrigation systems depends on the value ofthe incentive. But how can this bent be formalized in amodel or in an indicator so that the effects of a givenvalue of the incentive on the whole agriculture pro-duction can be assessed? Furthermore, the uncertainty ofthe physical processes that influence water demand (e.g.net radiation, canopy cover) couples with a behaviouraluncertainty (the proneness of farmers to change). Howcan such a mixture of uncertainties be described?

A solution to tackle with these two different uncer-tainties is to adopt Bayesian Belief Networks (BBNs)(Jensen, 1996, 2001). BBNs are directed graphical mod-els in which nodes represent random variables and thelack of arcs between two nodes represents the con-ditional independence of the two variables. Given twonodes A and B, one can regard an arc from A to B asindicating that A “causes” B. The model can be identifiedby specifying a Conditional Probability Table (CPT) at

each node. A CPT lists the probability that the child nodetakes on each of its different values for each combinationof values of its parents. BBNs can be effectively usedto model those components of the system for which theknowledge is limited or unstructured, and the cause–effect relationships are not evident, as in the case of theabove example. The insufficient prognoses in this typeof process would result in a weak description by simpledeterministic or markovian models. On the other hand,for the components of the system on which the knowl-edge is high and well structured (e.g. the hydrologic andhydraulic systems), the Bayesian approach would becumbersome and redundant, while the stochastic descrip-tion is perfectly suited (think for instance of the conser-vation of mass equation representing the reservoirdynamics).

To give an idea of how a BBN works, consider thenetwork of Fig. 5, which describes an irrigation district,such as the one of the Verbano example. Release, solarradiation, air temperature, soil state and biomass at timet (dark gray boxes) are the inputs of the model, whilethe financial incentive (squared box) is the planning vari-able; soil state and biomass at time t + 1 are the modeloutputs (light gray boxes). Inputs and planning variablesare called evidence variables, since given their value the(probabilities of the) values of all other variables of thenetwork may be inferred through evidence propagation.Since soil state and biomass appear both as input andoutput, they are the state variables of the BBN, and theBBN can be seen as nothing but a particular way ofwriting Eq. (3). Notice that the BBN in Fig. 1 containstwo distinct sub-models: the first describes the farmers’behavior, the second the plant growth. The farmers, viathe sector expert, are directly involved in identifying thefirst sub-model. They have to identify the items (parentsnodes) that could in some way influence their choices(already done in Fig. 5) and then express the prob-abilities of irrigating their fields and adopting a particularirrigation system conditioned to these items. The CPTsof the other sub-model may be filled either by interviewsor by using a classical compartmental model (e.g.CROPSYS, Caldwell and Hansen, 1993). The processof filling in the CPTs is conceptually analogous to thecalibration of a markovian model. It is now apparent thatBBNs are nothing but a generalization of markovianmodels (Smyth, 1998). In conclusion, BBNs appear tobe the right instrument to represent the relationshipbetween the water system variables and the indicators:when the relationship is of dynamical nature, Eqs. (3)are used together (and then ct is part of the water systemstate), while if the relationship is non-dynamical onlyEq. (3b) is active (since ct is non-existent). The first casehappens when the farmers’ indicator is the biomass atcrop time, the second when it is the supply deficit (seeFig. 6).


Fig. 5. A BBN dynamic model of an irrigation district.

Fig. 6. A BBN static model of an irrigation district.

9. Comparison and negotiation

Once all the alternatives generated have been evalu-ated, the SA can proceed to perform their comparison(phase 5). The most common approaches (e.g. Saaty,1980) determine the best compromise alternative in twosteps: first the stakeholders of each sector express theirpreferences among the sectors by means of weights, thenthe DM expresses his/her preferences among the stake-holders. In a way that depends on the adopted approach,these systems of preferences are merged into one, thusobtaining a vector of weights, one for each index. Bymeans of this vector the set of alternatives may beordered, ranking them from the most to the least desir-able. To adopt one of these approaches it is mandatorythat there exists one DM only, to whom the politicalpower of specifying the social relevance of the differentsectors is given, i.e. to whom the resolution of the con-flict among the stakeholders is delegated.

Often this is not the case, as it is not in the case ofthe Verbano system: since the system is a transnationalone, there are at least two DMs, the governments of Italyand of Switzerland. In this case, the best compromisealternative is the result of a negotiation process betweenthe two DMs. To supply the negotiation, the set of alter-natives has to be skimmed by identifying a set ofcompromise alternatives, each one preferred by a groupof stakeholders and possibly refused by others. Obvi-ously, when there is only one compromise alternative, ithas no opponents: therefore, it is also the best alternativeand the DMs have nothing to decide. On the contrary,it is the DMs’ task to determine the best compromise.To accomplish this task, they have to know not onlywhich are the alternatives of interest to the stakeholders,i.e. the set of the compromise alternatives, but also foreach alternative who are the supporters and who theopponents.

This information can be produced by establishing anegotiation process among the stakeholders that aims atidentifying alternatives with the highest consensus. Con-sensus does not imply complete agreement of the sup-porters, but means that each supporter feels reasonablycomfortable and accepts the alternative as a feasiblecompromise with other supporters. A procedure to ident-ify these alternatives can be derived from the so-calledPareto Race (Korhonen, 1988; Korhonen and Wallenius,1988). In its original version, it is a procedure devisedto help one single DM search for the best compromisepoint on a Pareto boundary, without determining thewhole boundary in advance. We do not have sufficientspace and time to describe the modified Pareto Race pro-cedure (see Soncini Sessa et al., in press). We may only


sketch the procedure by saying that the SA invites onestakeholder to indicate the alternative (s)he prefersamong all the alternatives. Then the SA shows, usinga diagram (Fig. 7), the utility level expressed by eachstakeholder for that particular alternative and invites thestakeholders who feel unsatisfied to specify the reason.The set of alternatives is then explored to find out a newalternative that improves the utilities of the unsatisfiedstakeholders, without lowering the utilities of the favor-able ones too much. If such an alternative does exist theprocedure is iteratively repeated, until it is not possibleto improve the situation of the worst of stakeholderswithout significantly reducing the utility level of theother stakeholders. This last alternative is a compromisealternative. By repeating the procedure for each one ofthe stakeholders, the set of the compromise alternativesis finally determined.

10. DSS architecture

We have already seen that the complex and recursivenature of the decisional process require a computer-based support system (DSS). In order to complete thepicture, we will shortly describe the main features ofits architecture, making reference to a prototype, namedTwoLe (Soncini-Sessa et al., 1999), that imbeds many(not yet all) of the ideas presented in this paper.

A traditional DSS gives access to data from anunstructured database; we think it is more appropriatethat the database architecture be tailored to obtain inde-pendence between data, models and processing algor-ithms. For that reason, the database of TwoLe has beenpartitioned into three different modules: the domain basethat contains the structural knowledge regarding data andtime series; the model base that contains the descriptionof mathematical models, both descriptive (simulationand forecast models) and prescriptive (decisional mod-els, i.e. multi-objective optimal control problems); the

Fig. 7. The alternative comparison.

experimental base that stores the formulation of modelidentification and/or policy design problems as well astheir results. In other words, data are stored following ahierarchical approach and manipulated according to anobject-oriented paradigm. Data in the domain and modelbases are classified into basic and compound objects.Building a basic object domain, e.g. a catchment or areservoir, is the first modelling step, where raw data areorganized by hypothesizing a modelling structure and itspurposes; then a set of basic objects can be grouped toconstitute a compound object, such as a water system ora distribution network. On the same object, either basicor compound, different models can be built: the purposeof the model base is in fact to promote model substi-tution and interconnection. The DSS/P and DSS/M usea common domain base and model base to perform theirtasks. Finally, the experiment metaphor has beenadopted to represent the SA activity: an experiment startsby choosing the necessary ingredients (data and/ormodels); continues by formulating a model identificationproblem or a multi-objective optimal control problem ora simulation; and ends by solving the posed problem andstoring the results. To compute the problem solution, atool (either an identifier or an optimizer or a simulator) isapplied, which can be picked up from a tool box. TwoLeprovides a set of standard identification tools, a widechoice of policy design optimizers and a bunch of simu-lators. The latter includes deterministic, markovian andMonte Carlo simulators, while the optimizers are basedon different algorithms such as stochastic dynamic pro-gramming (Bertsekas, 1995), neuro-dynamic program-ming (Bertsekas and Tsitsiklis, 1996; De Rigo et al.,2001) and Q-learning (Watkins and Dayan, 1992; Castel-letti et al., 2001). All these algorithms have been con-ceived to deal with set-valued policies.

The DSS control unit allows communication betweenthe two levels of the DSS and the domain, model andexperiment bases. In the present version of TwoLe, bymeans of the DSS/P, the SA can edit data, design, cali-


Fig. 8. The TwoLe GUI: calibration of an ARMA.

brate and validate models, produce optimal distributionpolicies (network optimization) and optimal release poli-cies (water system optimization), and evaluate the per-formance of alternatives by means of simulation. In otherwords, it supports all the activities required to completethe phases 2–4 in box A of Fig. 3. In the future, it willalso support phases 1 and 2 in box B. The DSS/M (boxC) can presently deal with forecasting, by using real-time data from a telemetering network, and compute thedaily release decision.

Finally, Twole has a friendly user interface (GUI) tohelp the SA in data preparation and algorithm choice.The GUI is inspired by the toolbox and folders metaphor(see Fig. 8): the SA can browse the folder structure asa normal file manager, and choose a tool to perform itstasks on the items stored in a folder. The folder structureis recursive: at the top level the items are meta-domainobjects; each one opens on real domain objects and, inturn, each domain contains its models and, for eachmodel, the experiments that were made on or with it.

11. Conclusion

In this paper, we have briefly reviewed the mainphases of the well known EIA procedure, pointing outsome challenging and significant concerns posed by itsapplication to planning and managing water reservoirsystems. For each one of these issues, we have proposedand analyzed some innovative solutions. In doing so, wehave stressed the key role played by a wider involvement

of stakeholders and DMs at every stage of decisionalprocess and suggested how this involvement could betechnically pursued. Finally, we have described somestructural features of TwoLe, a two-level DSS, whichimbeds some of (not yet all) the ideas presented. TwoLehas been successfully applied to the planning of the Ver-bano water system in Northern Italy (Betti et al., 2001)and is currently adopted in the recently started EU pro-ject MERIT (Management Environmental Resourcesusing Integrated Techniques) aimed at exploring theapplicability of BBNs as an integrated management tool.

Acknowledgements

The research was partly funded by the European Com-mission within the framework of the MERIT program,contract EVK1-CT-2000-00085.

References

Anthony, R., 1965. Planning and control systems: a framework foranalysis. Ph.D. Thesis, Harvard University Graduate School ofBusiness Administration.

Aufiero, A., Soncini-Sessa, R., Weber, E., 2001a. Set-valued controllaws in minmax control problem. In: IFAC Workshop Modellingand Control in Environmental Issues, August 22–23, Yokohama,Japan, Elsevier.

Aufiero, A., Soncini-Sessa, R., Weber, E., 2001b. Set-valued controllaws in TEV-DC control problem. Technical Report 2001.60, Dip.Elettronica e Informazione, Politecnico di Milano, Milan, Italy.


Bertsekas, D., 1995. Dynamic Programming and Optimal Control.Athena Scientific, Boston, MA.

Bertsekas, D., Tsitsiklis, J., 1996. Neuro-Dynamic Programming.Athena Scientific, Boston, MA.

Betti, E., Castelletti, E., Cellina, E., Passoni, E., Soncini Sessa, E.,Weber, E., 2001. A new regulation policy. In: Il Verbano: QualeFuturo? Insidie di un Ecosistema e Mezzi di Difesa, Ascona, Switz-erland, 9–10 November 2001, (in italian).

Brown, L., 2001. How water scarcity will shape the new century.Water Science and Technology 43 (4), 17–22.

Caldwell, R., Hansen, J. (Eds.), 1993. Systems Approaches to Agricul-tural Development, Simulation of multiple cropping systems withcropsys. Kluwer Academic Publishers, Dordrecht, The Nether-lands, pp. 397–412.

Castelletti, A., Corani, G., Rizzoli, A., Soncini-Sessa, R., Weber, E.2001. IFAC Workshop Modelling and Control in EnvironmentalIssues, August 22-23, Yokohama, Japan, A reinforcement learningapproach for the operational management of a water system. Elsev-ier.

De Rigo, D., Rizzoli, A., Soncini-Sessa, R., Weber, E., Zenesi, P.,2001. Neuro-dynamic programming for the efficient managementof reservoir networks. In: MODSIM International Congress onModelling and Simulation, The Australian National University,Canberra, Australia, December 10–13,

Jensen, F., 1996. An Introduction to Bayesian Networks. Springer-Ver-lag.

Jensen, F., 2001. Bayesian Networks and Decision Graphs. SpringerVerlag.

Keeney, R., Raiffa, H., 1976. Decision with multiple objectives: Pref-erences and Value Trade-offs. John Wiley & Sons, New York, NY.

Korhonen, P., 1988. A visual reference direction approach to solvingdiscrete multiple criteria problems. European Journal of Oper-ational Research 34, 152–159.

Korhonen, P., Wallenius, J., 1988. A pareto race. Naval ResearchLogistics 35, 615–623.

Ljung, L., 1987. System Identification: Theory for the User. Pren-ticeHall, Englewood Cliffs, NJ.

Loucks, D., 1990. IRIS: An Interactive River System Simulation Pro-gram—User’s Manual. IIASA, Laxenburg, Austria.

Loucks, D., Kindler, J., Fedra, K., 1985. Interactive water resourcemodeling and model use: an overview. Water Resource Research210 (2), 95–102.

Saaty, T., 1980. The Analytic Hierarchy Process. McGraw Hill, NewYork, NY.

SFWMG, S. F. W. M. D. OASIS, the South Florida Water Manage-ment District#s operations artificial intelligence program. DistrictBrochure, 1987.

Simonovic, S., Savic, D., 1989. Intelligent decision support and reser-voir management and operation. Journal Computing in Civil Engin-eering 30 (4), 367–385.

Smyth, P., 1998. Belief networks hidden Markov models, and Markovrandom fields a unifyng view. In Pattern Recognition Letters.

Soncini Sessa, R., Betti, E., Cellina, F., Passoni, A., Weber, E. TheVerbano Project. McGraw-Hill, Milan, Italy, in press (in italian).

Soncini-Sessa, R., Canuti, D., Colorni, A., Laniado, E., Losa, F., Riz-zoli, A., Villa, L., Vitali, B., Weber, E., 2000. Use of multi-criteriaanalysis to resolve conflicts in the operation of a transnationalmultipurpose water system, the case of Lake Verbano (Italy–Switzerland). Water International 250 (3), 334–346.

Soncini-Sessa, R., Rizzoli, A., Villa, L., Weber, E., 1999. Twole: asoftware tool for planning and management of water reservoir net-works. Hydrological Sciences Journal 44 (4), 619–631.

Soncini-Sessa, R., Zuleta, J., Piccardi, C., 1991. Remarks on the appli-cation of a risk-averse approach to the management of “El-Carri-zal” reservoir. Advances in Water Resources 130 (2), 76–84.

Watkins, C., Dayan, P., 1992. Q-learning. Machine Learning 8,279–292.

Yakowitz, S., 1982. Dynamic programming applications in waterresources. Water Resources Research 18 (4), 673–696.

A DSS for planning and managing water reservoir systems

Documents

Transcript of A DSS for planning and managing water reservoir systems