-

PROBABILISTIC CHOICE VS. RISK ANALYSIS CONFLICTS AND SYNTHESIS

IN POWER ~- SYSTEM
-~ PLANNING
Vladirmro Miranda L. Miguel Proenqa
INESC-Macau - Inst. Eng. de Sist. e Computadores de Macau, and INESC - Instituto de Engenharia de Sistemas e Computadores,
FST- Universidade de Macau, MACAU PORTUGAL
emnl: [email protected] Fax: 35 1.2.2084172

Abstract - This paper shows the conceptual differences 2. THE PROBABILISTIC CHOICE PARADIGM
between adopting a probabilistic weighting of the futures and a
risk averse strategy, in power system planning under uncertain One of the most successful ways of modeling uncertainties has
scenarios. It is illustrated with a distribution planning problem, probabilities at its core. In power system planning, there is a
where optimal solutions in both cases are determined by a vast acquired patrimony of knowledge and methods that use
Genetic Algorithm. It shows that the probabilistic approach is them, from reliability to stochastic inflow in hydro system
less safe and cannot detect some interesting solutions. models, just to give two examples.
The justification for the adoption of a probabilistic model must
1. INTRODUCTION be found on the basic concepts of probabilities themselves. The
underlying assumption is that there is a non-variant set of laws
Many mathematical models have been proposed in the past for goveming the fixed conditions for a repetition of
optimal electrical network planning, but the vast majority had experiments whose outcome is random in nature.
been, until recently, aiming at a so called optimal solution
and were based on single criterion optimization. These fxed experiment conditions may be observed or
simulated on successive events on a single element of a model
The authors have adopted a new approach to the power system (such as the water inflows in a river basin throughout 30 years)
planning problem. We have demonstrated that it was possible or on a family of similar elements, under similar experiment
to use Genetic Algorithms (GA)[I] to find optimal dynamic conditions (as it is assumed to happen to network components,
expansion planning solutions[2], how the GA could be used to in order to form a basis for a failure-repair model).
define an expansion strategy within a tree of futures[3] and that
a continuously fuzzy modeling of each future could be When, in planning problems, one allocates probabilities to
combined with the scenario technique, under the form of a tree different possible futures, one is keeping this basic assumption
of hzzy futures, within a GA, to generate an expansion - that the laws goveming the phenomena remain unchanged.
strategy under uncertainty[4,5]. Furthermore, another assumption exists: that there will be a
significative number of experiments, so that the frequency of
The relevant aspect to note is that the use of GA made it very occurrence of whatever phenomenon one is considering will
easy to have the new models following the concepts of Risk tend to be close to the probability value assigned to it.
Analysis (RA) and therefore favoring solutions that minimized
regret (a distance between the chosen solution and the optimal) The Probabilistic Choice paradigm is therefore the following:
instead of solutions that minimized the cost (whatever this may admit a planning problem and a cost function to measure the
be) on the average of the futures, using what we will call the goodness of a solution (latu sensu: we refer as cost any
Probabilistic Choice paradigm (PC). criterion that may be considered and adequately transformed
into an objective of minimization); given a set of futures, each
The PC paradigm is well known and is the basis of a vast one with a probability value assigned, the optimal solution
number of planning models and techniques. The RA paradigm should be chosen among those that minimize the expected cost
is certainly much less widespread; very strong arguments in over the set of futures considered.
favor of this perspective have been aligned in [6] and since
then the recognition of its merits has been growing at a steady In some cases, although no real probabilities can be assigned to
pace. In other scientific areas, these principles have also been each future, planners still define weights in relation to them, as
aplied (check [7], for both PC and RA approaches). References a way of expressing their perceived relative likelihood. The PC
[8,9,10,11,12,13] illustrate examples of the growing interest in paradigm still holds, however, as long as the preferred solution
Risk Analysis principles - however, any literature review must is chosen as stated before.
be done carefully because authors in many occasions used
loosely words like risk, robustness and other, expressing The PC paradigm is an a priori evaluation: the solution is
different concepts. One may find e.g. the word risk chosen before knowing which future scenario OCCUTS, and
associated with the variance of solution results (like in [14]), before evaluating any possible discrepancies between the
which is not what we address here. scenario that actually occurs and the weighted average of
futures accepted. If the cost incurred in scenario k is fk and the
This paper is devoted to analyze the differences between both probability of this scenario is wkthen the PC paradigm may be
PC and RA paradigms and the consequences of adopting either characterized by (where the index i stands for any possible
one in power system planning. solution):

3. THE RISK ANALYSIS PARADIGM

Under the Risk Analysis paradigm, the preferred solution is
one that minimizes the regret felt by a Decision Maker (DM),
after verifying that the decisions he had made before, were not
the optimal ones given the future that in fact occurred.

0-7803-3713-1/97 $10.00 0 1997 IEEE

16
An easy measure of this regret may be defined as the difference 4. CONCEPTUAL DISCUSSION
between the cost of the adopted solution, and the cost of an
optmal solution that would have been adopted, if the DM Is one paradigm more adjusted or adequate than the other? Is
knew beforehand which of the future scenarios would happen. there place for both, in selected and distinct planning
problems?
The concept of hedging plays an important role in the RA
paradigm: in short, hedging means that a DM will opt for a From the discussion above, we see that the PC paradigm must
more costly solution, if this reduces the negative consequences hold whenever we can be reasonably sure that the problem is
of possible adverse futures, instead of choosing a cheaper within a frame of repetition of events. In this case, we may be
solution that, however good in most futures, may lead to a statistically sure that bad decisions at one time will be
heavy loss in a particular scenario. compensated by very good decisions at a later time, so that
pickmg up the optimal average must be the right strategy to
In this context, it is usual to adopt the designation robustness to obtain an optimal trajectory along m e .
express the characteristic of a system design to be accepted as
good for a sub-set of future scenarios; one alternative may be We can easily show that the minimization of the average regret
said as 100% robust if it would be good in all scenarios. The felt for not choosing the optimum is equivalent to the PC
concept of exposure comes closely linked with robustness. paradigm, symbolized by equation 1:
Exposure denotes the fact that an alternative is not acceptable
in some sub-set of futures. Its degree may be measured just in
terms of a number of unacceptatble futures, or then on some
kind of distance from a level that would make it acceptable.
Within the RA paradigm, the scenarios been taken as ' k
admissible futures may also be weighted, according to their
perceived likelihood by the DML However, this must not be this being so because the term c W k f i p is a constant and
done lightly: in the RA context, the DM is trying to perceive k
the risks he is running for taking some decision, and therefore therefore does not interfere in the process of finding an optimal
it is important not to mask catasbrophic events with low weight rmnimizing solution.
values (at least in the first stages of the planning process,
before the DM becomes aware of them). It must become clear by now that unless our planning model
accounts for the repetition (in statistically significant large
To define a systematic method to determine a preferred number) of the scenarios and the decisions taken, it is not
solution, under the RA paradigm, we may first admit that we possible to consider the PC paradigm as adequate - the
have a way of calculating, for each scenario, the optimal frequency will deviate from the probability of occurrence and
solution and its cost fkOp, assuring that we know for certain the result obtained will deviate fiom the expected value.
that this is the future that will occur (a deterministic context, no One example of a problem where the PC paradigm seems
uncertainty). Then we must have a way of calculating, for each valid the mid-term operation planning of a hydro system. It is
alternative i in future k, the Regret felt for not having chosen reasonable to assume that the water inflows will be adequately
the best for that future. The RA paradigm may be now modeled by a stochastic model, and the time frame considered
characterized by may be large enough to allow enough repetition of situations
and decisions taken, so that deviations between the average
future and the actual inflows will get compensated on the long
run. Of course, to be more strict we are assuming that the
system will remain with the same components, that the climate
This means that we search, in each solution i, for the maximum and water inflows will not change, etc.
regret among all futures k, and then search among all solutions One example where the PC paradigm seems to be not so
i to look for the one with the smaklest maximum regret. adjusted: the long range expansion planning of a power system.
The RA paradigm is an a posteriori evaluation: the solution is It is difficult to see how the decision to build a given power
chosen after assessing the consequences, in each admitted station, if proven to have been inconvenient, is going to be
possible future, of the decisions taken; the choice is based on statistically compensated in the future. In fact, one single
the consequences, and not on the solutions themselves. The unadjusted decision may even prove to be catastrophic,
keyword here is decisions. The RA paradigm changes the focus disrupting all eventual compensation process. For a case such
from solutions to decisions. as this, the RA paradigm seems to be a more adequate answer,
in t e r m of methodology.
The RA approach is more intixnately related with Decision
Making theory than the PC paradigm. Because it evaluates
regrets, the preferences and risk aversion of the DM must be 5. A MULTICRITERIA MODEL FOR THE "PLANNING
taken in account in the process of selecting acceptable UNDER UNCERTAINTY" PROBLEM
solutions. The formulation of e:quation 2 is general; but a
simplification may be acceptable, in those cases where it can be We will illustrate now some further differences between the PC
argued that the regret felt by the IDM is just proportional to the and the RA paradigms, by adopting a multicriteria model for
cost deviation to the optimum: the planning problem under uncertainty represented by
scenarios.
Admit that in a planning problem we have only two likely
(3) scenarios 1 and 2, and that we have available only four
solutions A, B, C and D to consider. For each solution we have
This most likely happens in problems where alternative evaluated the costs in each scenario , which led to evaluating
solutions display a narrow range for objective function values regrets, such as given in the following table (points shown in
and no catastrophic scenarios are foreseen. Figure 1).

17
 Scenario 1
10
8
6
4
2
0
0 2 4 6 8 1 0
Scenario 2
D
models
R4-

- ED I
PC -
0
A B

05
C

weight w l of future 1
D

Figure 2 - Stability ranges of wl in both PC and RA paradigms The

ranges obtained with an Euclidean Distance model E D are also
represented
Solution A Solution B Solution C Solution D
Scen. 1 8 6 3 0 The PC paradigm, with its L, metnc, a linear composition of
Scen. 2 0 4 5 10 the ObjeChVeS, can just trace the convex hull of the non-
dormnated set and therefore is blmd to solution B The RA
Applying equations 3 and 4, we can obtain the preferred paradigm, usmg a non-hear metric where the isovalue llnes
solutions, under the RA and PC paradigms. For instance, if we form a nght angle over any point (x=y), after rescaling axes,
assign the weights (or probabilities) to each scenario, such as can explore concave areas of the non-domiated border and
w,=O.7 and w,=0.3, we obtain Solution D for PC approach, and therefore find other mtereshng solutions, llke solution B in our
solution C for RA approach. This immediately shows (which example.
should not be surprising) that the two paradigms may lead to Another unportant fact becomes evident: the PC approach
different proposals. tends to favor extreme sohhons, while the RA approach has a
Much more interesting is to verify the weight stability, i.e., more cauaous and balanced sweepmg of the non-donated
within which ranges of assigned probabilities to the future soluhons. We can see in figure 2 that soluhons A or D are the
scenarios are the best proposals stable. The confidence in a preferred SOhhOnS for larger ranges of w, values m the PC than
solution or a decision must be built also upon verifying if one m the RA approach - but these are precisely the nskier
is allowed some margin of error in assessing scenario sohhons (they have associated the larger regret values)!
probabilities. The RA approach keeps, on a larger range of probability
Figure 2 gives us the answer for the example problem we have hypothesis, the preferred solutlon i the knee region of the
been following (the bar labeled ED will be explained later). non-dormnated border. Soluhon C alone is recommended under
RA for a much broader range of weights than under PC (not to
For example, solution A would be preferred in scenarios menhon B and C together).
defined by w, E [O; 0.51, under the PC paradigm; however, for I
an important range of weights, solutions B or C are less risky. We draw, therefore, these unportant conclusions.
Before furthering additional comments, lets call the readers * the PCparadigm results in models which may be 1
attention to the kind of analysis we are performing. In fact, we blind to solutions that are interesting to be
are following all the procedures usual in multicriteria decision considered in an uncertain environment
making, for a problem with two objectives - and that is the
model that we propose, to consider the planning problem in * the PC paradigm tends to recommend, in many
multiple scenarios as a multiobjective one, each objective being cases, riskier decisions
the minimization of the regret in each scenario. It is mtereshng to nohce that Zeleny defied compromise set as
In our example and in general, the origin (where all regrets are the set of soluhons close to the ideal, that would be obtamed by
zero) is precisely the Ideal solution, as proposed by Zeleny varymg the metnc used to measure this distance, namely from ,
[ 151 - a virtual solution that would be chosen by the DM if it L, to L,. The calculahon of the compromse set is a way of
were feasible (it is the best in all objectives). Under this reduclng the dunension of the set of altematwes.
multicriteria perspective, all solutions A, ..., D are non- It is appealing to suggest that a compromse set among
dominated and therefore natural candidates to be the final scenanos be calculated in the power system plannmg problem.
choice by the DM [we recall that a solution is called non-
dominated, or Pareto optimal if, for all remaining solutions, It is also tempting to consider an Euclidean metnc L, as a sort 1
having a better performance in whatever criterion means that it of possible compromise between the PC and the RA paradigms. ~

must be worse in at least one other criterion].

Also under this perspective, the PC and the RA receive the
following interpretation:
* the PC paradigm, translated by equation 4, is equivalent to
fiiding the point closest to the Ideal, using a L, metric; Figure 2 shows us what happens with an Euclidean distance
0 the RA paradigm, translated by equation 3, is equivalent to metric, in the bar ED. It is clearly an intermediate step between
finding the point closest to the ideal, using a L, metric. PC and RA, because it favors less the extreme risky solutions, ~

but not to the same extent s the RA approach does it.

The weights (or probabilities) wk just represent a re-scaling of
the objective function values. However, the ED model in not enough non-linear to be able
to select solution B as a preferred one for a range of weight
Under this light, we can better understand what happens in the values - in this example, the concave part of the non-dominated
example we have been following. border remains hidden and is not detected by the ED probe.
6. A DISTRIBUTION PLANNING EXAMPLE In any case, for the example worked out for this paper, we have
defined a trade off value f for the MW repressed, such that
6.1. Regrets
In order to reinforce the conchions above, we now present a
more complex example, worked out as a distribution planning
problem. The methodology adopted for the solution such a This represents how much one is willing to Invest (Inv) to
problem was the one described in [4,5], which includes the reduce the repressed demand (Pq) in the amount of 1 MW.
representation of uncertainties by a tree of fuzzy futures. But to
avoid introducing too much commplication that could obscure Having established a reference value f for planning purposes,
the interpretation of results, we have considered no fuzziness in an optimal solution combining investments and repressed
data and introduced other simpl~ifications,not relevant to the demand may be given by
discussion here.
min Inv + T P,,
We will defme an expansion planning problem in three time
stages, with three scenarios for load growth, forming a crisp This definition is important for understanding the next
tree of futures. The methodology for obtaining a compromise paragraphs, because we need to define Regret in this context.
planning solution, considering the uncertainty about the future, A regret measure, for a solution in each scenario, will be
is summarized as follows:
conditioned by the following reasoning:
Consider each path in ,the tree of futures as 100%
certain; define an optimization problem for this path: a) if the investment is excessive relative to the optimal
the solution will constitute a path ideal for this investment in that scenario (given by the path ideal), then the
scenario. regret value will be basically the difference between the money
actually spent and the one that would have been spent if that
Solve the dynamic planning problem for each path using scenario were known as certain (the cost of the path ideal);
a Genetic Algorithm. b) if the investment is insufficient to supply all forecasted load,
Define now a compromise objective, considering all then the regret will be basically the economic equivalent of the
scenarios or paths in the tree offutures as possible: this repressed demand originated.
compromise objective must translate the idea that we
seek for solutions as close as possible to the path Of course, in practice, one calculates the total (investment +
ideals previously found, or that we wish to avoid being repressed demand) cost for each solution, and derives the regret
too farfrom these path ideals (which is not the same values from the comparison with the path ideal.
thing).
6.2. Worked example
Solve the new problem with a Genetic Algorithm (GA).
Consider now a distribution planning problem for the network
Inllref. [2,4,5] one may fmd descriptions of the type of GA in Figure 3 - Existing network, possible new substation,
adopted and the coding tactics adopted. The authors have composed by 54 nodes, 16 lines and two substations in the
developed a very efficient parallel processing GA platform, initial system and 45 lines and 2 substations in project: solid
running on a cluster of workstations and PCs, which allows lines represent the initial system; new lines may be built to
obtaining results with short computing and total elapsed times. connect new load points.
The model described in previous publications is very complete; To keep it simple, we called the departing system as Time
besides admitting fuzzy data, it also allows taking in Stage A and divided the planning horizon in two more Stages,
consideration several criteria to select the preferred solutions: B and C.
minimizing investments or power losses, maximizing voltage
quality and reliability, minimizing system exposure and system
inadequacy (by minimizing the possible repressed demand due
to insufficient planned capacity compared with actual load
growth).
In order to keep it simple, we decided in the following example
to deal mainly with costs, of two natures:

:i investment
repressed demand
We defined repressed demand as the difference between s3
planned system capacity and forecasted load. In distribution
lines, if there is insufficient capacity, the utility will not allow
more consumers to be connected, generating delays in
satisfying requests for increased supply (and originating added
costs in unplanned supply alternatives or remforcements).
These delays have a social and economical cost that must be
accounted for. These are, in na.ture, different from the ones
resulting from load disconnectioins following contingencies, as 10Km
I

calculated in reliability studies; and we must agree that there Figure 3 - Existing network, possible new substation sites and
are not many indications on how to evaluate one kW of location for forecasted new loads. Thick lines represent
repressed demand - except that it might not be correct to existing branches and thin lines represent possible sites for the
associate to it the same economical value as estimated for the expansion of the system. Substations S3 and S4 are considered
kW disconnected. to be in project.

19
All data were considered crisp. The only uncertain factor was corresponds to a very real engineenng practice, and could have
load forecast, for which a tree of futures has been defined, such followed a number of methods). For our purpose, the network
as in Figure 4, establishing three Paths or Scenarios labeled P1, designs are not important; but their regrets in each scenano are
P2 and P3. The total area loads for each scenario are the following:
(MW) I s I stageB 1 Stage C
Scenario P1 I 404 I AX7 1 57 0
ScenanoP2 I 404 I 523 I 68 1
ScenanoP3 1 40.4 I 570 I 79 5

Substation sizes and other complementary system data are

shown below and 111 the next table
.. ... ..
Feeder and substation cost crisp 11177.3 I 0 I 41 3 I 0 I 0 I 1.3 I 177.3 141.3 I 26
Discount Rate 10%
Nominal voltage (W 15
Voltage drop threshold 8Yo In a scenario space [Pl, P2, P3], the ideal solution in the
Investment/r.demand T ($/MW) 20 sense of Zeleny is the point (O,O,O) - no regret in any future.
We can apply expressions (3), (4) and ( 5 ) to obtain the solution
which is closest to the ideal, according to RA, PC and ED
distance models - but we need a defmition of the values w, the
weights or probabilities assigned to each scenario.
If these are taken as parameters, we can represent the closest
to the ideal solution in a Space of Weights. wl, w2, w3,
The definition of the problem shows us that at Stage A one showing for which combination of weights is each solution
strategic decision must be made, in a compromise for the remaining preferred.
unknown development of demand. Once this decision having As wl+w2+w3=1, we can project this result in a plane [wl;
been taken, there is a dynamic planning problem for each Path w2]. This exercise has been done for the PC, ED and RA
but no uncertainty. models and its results are depicted in Figure 5 , in a
representation in the space of weights [7].
P3
The comparison between PC and RA results is again very
P2 enlightening: the PC model suggests the extreme solutions S 1,
PI 52 and S3 for much larger ranges of probability values than the
RA model. Furthermore, the PC approach just ignores the
compromise alternative S4, which would be chosen by risk
35 time averse Decision Makers for a reasonably important band of an
A B C assessment of the likelihood of the three scenarios.
Figure 4 - Tree offutures as afunction of loadforecast The ED model helps us to understand what happens when we
change of point of view, from PC to RA. Solution S4 is
revealed and the area of preference of S1, S2 and S3 starts
This path problem is exactly the one to be solved according to shrinking.
point 2 of the algorithm in the previous section, using the
method described in [2].
Three path ideals have therefore been calculated, labeled w2 w2
I
solutions S1, S2 and S3 with investment costs of 753, 889 and
971 x IO6 PTE (Escudos). In each path ideal, of course, no oa
repressed demand exists. s5 06
Having these anchor points calculated, phases 3 and 4 of the 04
algorithm in previous section could be executed. We have
02
applied again a GA, similar to the one in [ 5 ] , which helped us
in finding solutions located on the non-dominated region in a W1
0
multi-criteria problem that could be defined through , 0 0.2 0.4 0.6 0.8 1 0 0.2 04 06 0.8 1
minimizing three criteria:
WZ

1
1
min Regret( P1)
OB
min Regret(P2) 06

min Regret(P3) 04

02
A number of interesting non-dominated solutions have been 0
found. Because this is a worked example, we have simulated a 0 02 0 4 06 08 1 wl
decision process of analysis of each alternative and have kept 5
solutions (this is somewhat arbitrary, but the decision process Figure 5 - Preferred solutions in the space of weights,for PC,ED and
that philters the set of solutions and leaves only five RA models. I

20
In any case, these results may .also be seen as reassuring: in
many combinations of scenario probability values, the REFERENCES
solutions recommended by the PC and the RA approach are the D. E. Goldberg, Genetic algorithms in Search, Optimization
same. The lack of conflict in these cases may serve as a and Machine Learning (book), 1989 Addison-Wesley
confirmation that it would be difficult to argue that a better
decision could have been taken! V. Miranda, J. V. Ranito, L. M. Proenqa, Genetic Algorithms
in Optimal Multistage Distribution Network Planning,
7. CONCLUSIONS IEEERES 1994 Winter Meeting, New York, New York, Jan.
94, in IEEE Trans. on PWRS, Nov 94, pp. 1927-1933
This paper was devoted to analyzing the assumptions behind
known planning paradigms, labeled as the Probabilistic Choice L.M. Proenqa, V. Miranda, Genetic Algorithms in Optimal
Electric and Gas Distribution Network Planning,
(PC) and the Risk Analysis (RA) approaches, and evaluating Investigaq2o Operacional (Operations Research) review, voi
the consequences of adopting either point of view. 14, n.2, December 1994
The results from the examples presented speak for themselves: V. Miranda, Power System Planning and Fuzzy Sets:
the PC paradigm may lead to adopting riskier decisions than Towards a Comprehensive Model Including all Types of
the RA paradigm, and because it is a linear criterion, it remains Uncertainties, invited paper, 4th PMAPS - Probab. Methods
blind to a number of compromise solutions that do not Applied to Power Systems, Rio de Janeiro, Brazil, Set 1994
happen to lie on the convex hull surface of the non-dominated V. Miranda, L.M. Proenqa, A General Methodology For
border among scenarios. Distribution Planning Under Uncertainty, Including Genetic
Although the examples presented were simple for obvious Algorithms and F u u y Models in a Multi-Criteria
reasons, the conclusions are independent of the size of the Environment, Proceedings of Stockholm Power Tech -
problems, because they derive both from the comparison SPT95 Conference, Stockholm, Sweden, Jul95, pp. 832-837
between paradigm assumptions and their mathematical H.M. Meml, A.J. Wood Risk and Uncertainty in Power
structure. System Planning, i0th Power System Computation
Conference - PSCC, Graz, Austria, Aug 90
When adopting the PC paradigm in planning, one must
therefore check if its basic assumption (the repetition of events R. Clemen, Making Hard Decisions, PWS - Kent Publishing
introducing an effect of compensation) is verified. Company, Boston, 1991
The definition of scenarios, as a technique to represent E. Crousillat, P. Dorfner, P. Alvarado, H. Merril, Conflicting
uncertainty, demands from the Decision Maker the setting of Objectives and Risk in Power System Planning, IEEE
weights or probabilities for each scenario, expressing his Transactions on Power Systems, Aug. 1993, 8(3), pp. 887-893
sense of likelihood about each. These are uncertain values in B. Gorentsin, N. Campodonico, J. Costa, M. Pereira, Power
themselves (although not of a probabilistic type!), and System Expansion Planning Under Uncertainty, IEEE
therefore it is important to check the robustness of decisions as Transactions on Power Systems, Feb. 1993, pp. 129-I36
a function of changing weights.
B.G. Gorenstin, N.M. Campodonico, J.P. Costa, M.V.F.
The work presented shows that, in many cases, the solutions Pereira, Power System Expansion Planning Under
proposed by PC and RA will be the same. A sound planning Uncertainty, 1 Ith PSCC, Avignon, France, Set 1993, pp.
methodology should therefore look for this type of 585-592
confirmation. H.L. Willis, Spatial Electric Load Forecasting (book), pag.
A new model, called Euclidean Distance (ED) has helped in 14-17, ed. Marcel Dekker Inc., NY,1996
gaining insight on what happens when changing of point of C. J. Andrews, Evaluating Risk Management Strategies in
view, from PC to RA. We have shown that ED may also in Resource Planning, IEEE Transactions on Power Systems,
some cases reveal solutions ignored under the PC paradigm. Feb. 1995, Vol IO, Number I, pp. 420-426
Most of all, if hesitating (surely Cor sound reasons) between the
two approaches, the ED model may just give a compromise EPRI, RISKMIN: An Approach to Risk Evaluation in
approach between a probabilistic sense and risk aversion. Electric Resource Planning, rp. EL-5851, Palo Alto CA, 1988
In any case, and with the understanding that each planning R. Tanabe, K. Yasuda, R. Yokohama, H. Sasaki, Flexible
Generation Mix Under Multi Objectives and Uncertainties,
problem has specific features, this work comes certainly in IEEE Transactions on Power Systems, May 1993, pp. 58 1-587
support of an a posteriori view of decisions, contrary to the
traditional a priori assessment of solutions, and challenges M. Zeleny, Multiple Criteria Decision Making (book), Mc.
power system scientists and planners to pay more attention to a Graw Hill ed., 1982
Risk Analysis approach in the development and use of planning Vladimiro Miranda was bom in Porto, Portugal, on March I I , 1955.
models. He received his Licenciado, Ph.D. and Agregado degrees from the
Faculty of Eng., University of Potto (FEUP) in 1977, 1982 and 1991,
in Electrical Engineering. In 1981 he joined FEUP. In 1985 he joined
also INESC - a research Institute for Engineering in Systems and
Computers, held the position of Head of Power Systems area and was
responsible for several research projects in Portugal and within
European Union programs. Since Feb 1996 he is in the University of
Macau (China, under Portuguese administration) as Full Professor. He
also is the Chairman of the Executive Board of INESC-Macau.
AKNOWLEDGEMENT
Luis Mime1 Proenca was bom in Lisbon, Portugal, on May IO, 1966.
The work described on this paper was partially financed by He received his Licenciado and MSc. degrees from FEUP in 1989 and
JNICT (Project PRAXIS 3/3. I/TIT/43/94). 1994, in Electrical Engineering and Computers. In 1990 he joined
INESC as a researcher in the IDEIA area. He is now a Ph.D. student.

21