Professional Documents
Culture Documents
International Journal of Innovations in Energy Systems and Power (Vol. 4 no. 2, October 2009) 22 of 34
J for the hydrothermal coordination problem. In this work, a
Gt = ∑ g j ,t (7) penalty function approach [3] has been used to reduce the
j =1
chances of producing infeasible elements, but still allowing
g min ≤ g j ,t ≤ g max (8)
j j search trajectories from the outside of the feasible region,
I
Pt = ∑ p i ,t (9) which usually facilitate the localization of optimal solutions.
i =1 Therefore, constraints violations have been included in the
⎛ ⎞ Δt fitness function with penalty factors of the order of the
xi ,t = xi ,t −1 + ⎜⎜ yi ,t + ∑ uk ,t − ui ,t ⎟⎟ ⋅ 6t (10) expected solution.
⎝ k ∈Ωi ⎠ 10
hi ,t = φ (xi ,t ) − θ (ui ,t ) − pci ,t
avg
(11) B. Particle Swarm
xi ,t −1 + xi ,t Literature shows promising results from the application of
,t =
xiavg (12) PS to operational planning problems [4]. However, only the
2 thermal unit commitment problem has been studied in detail,
pi ,t = ki hi ,t qi ,t (13) and the potential of PS for providing hydrothermal
ui ,t = qi ,t + vi ,t (14) coordination solutions still needs investigation. Reference [5]
has established a general calibration for the PS search
xmin
i ,t ≤ xi ,t ≤ x max
i ,t (15)
(control) parameters. This calibration has also been proved
umin
i ,t ≤ ui ,t ≤ u max
i ,t (16) appropriate for power system problems [6].
In the present work, a different idea is employed. The PS
, t ≤ qi , t ≤ qi , t ( hi , t )
qimin max
(17)
method has a variant called Constriction Factor Approach [7],
vi ,t ≥ 0 (18) which has improved the robustness of the search. The
1 constriction factor χ is defined as a function of k, ϕ1 and ϕ 2 ,
λt = (19)
(1 + r ) where k∈[0;1]. Then,
t
⎣ ( ) ( )
vik +1 = χ ⎡vik + c1rand1 pbesti − sik + c2 rand 2 gbest − sik ⎤ (20)
⎦
where, for each thermal unit j and each hydro unit i
T : number of intervals; 2k
χ= (21)
I : number of hydroelectric plants; 2 − ϕ − ϕ 2 − 4ϕ
J : number of thermal plants;
λt : discount factor for interval t; where
r : discount rate for interval t; ϕ = c1+c2, for ϕ >4;
ki : productivity of plant i [MW/((m3/s).m)]; vik +1 : velocity of particle i at iteration k+1;
pi,t : hydro unit i generation during interval t; c1, c2 : weight factors;
Gt : total thermal generation during interval t; rand : uniformily distributed variable ∈ [0;1];
Pt : total hydro generation during interval t; sik : ith particle position at kth iteration;
Dt : load level during interval t; pbest : best so far obtained by the ith particle;
xavgi,t : average volume for unit i during interval t; gbest : best so far obtained by any particle; and
hi,t : average water head for unit i during interval t; k : exploration / intensification parameter.
ui,t : total discharge flow rate for unit i during interval t;
yi,t : inflow rate for unit i during interval t; The constriction factor χ allows the tuning of the
Δtt : length of interval t (one month); and exploration capacity by varying k. When k is close to zero, the
Ωi : set of upstream hydro plants. swarm is not allowed to explore distant regions of the search
space, which is convenient when promising valleys have
II. METAHEURISTICS FOR OPTIMIZATION already been found. On the other hand, for k close to unit, the
Four metaheuristics are used to solve such a non-convex particles are “free” to look for distant promising valleys.
optimization problem for a Brazilian real system. Their Finally,
implementation details are described next, along with all data
specification. sik +1 = sik + vik +1 (22)
where
A. Genetic Algorithms sik+1 : new position for particle i;
In large-scale optimization problems and in hydrothermal sik : previous position for particle i.
coordination in particular [1,2], real coding has been
successfully adopted in GAs because of its simplicity and the
C. Simulated Annealing
advantage of facilitating the definition of special purpose The search parameters for SA are:
search operators. • initial temperature T0;
Procedures for avoiding infeasible elements in the mating • annealing schedule given by Tk+1 = g(Tk); and
pool may not be appropriate in case of optimum solution on • number of transitions Nk for a Tk.
the boundaries of the feasible region, which is usually the case
International Journal of Innovations in Energy Systems and Power (Vol. 4 no. 2, October 2009) 23 of 34
There are some proposals in the literature to determine a Maximum
convenient initial temperature [8]. In this work, the following Plant Generation [R$/MWh]
procedure has been successfully adopted: [MW]
+ (1) Pernabuco 638 60.00
ΔV (2) Fortaleza 347 66.74
T0 = (23)
⎛ 1 ⎞ (3) Fafen 151 71.29
ln ⎜ ⎟ (4) Ceará 220 82.72
⎝ X0 ⎠ (5) Bahia 186 87.12
+
where ΔV is the average degradation on the objective (6) Camaçari 347 130.50
Total: 1889 ---
function values, with respect to the current solutions, for
Deficit Cost --- 855.31
transitions under the initial temperature that do not improve
their current solutions. X0 denotes the fraction of acceptance Table 1: Thermal generation capacities and costs.
for those transitions, which is usually made equal to 0.85.
The number of transitions Nk for each temperature level is Thermal
defined in this paper as a constant value. The temperature Generation Commitment Cost [R$/h]
scheduling has been implemented as follows: [MW]
Tk +1 = β Tk (24) 0≤ Gt≤ 638 (1) 60 Gt
638< Gt≤ 985 (1), (2) 66.74 Gt –
for β < 1 . 4300.12
985< Gt≤ 1136 (1), (2), (3) 71.29 Gt –
8781.87
D. Evolution Strategies 1136< Gt≤ 1356 (1), (2), (3), (4) 82.72 Gt –
Real-valued vectors have been used, following the same 21766.35
representation scheme applied by the other metaheuristics. 1356< Gt≤ 1542 (1), (2), (3), (4), (5) 87.12 Gt –
Recombination of pairs of parents, selected via tournament, is 27732.75
implemented by creating offspring with values taken from one 1542< Gt≤ 1889 All thermal units 130.50 Gt –
94624.71
of the parents (with equal likelihood). Survivor selection uses
Gt> 1889 All thermal units plus 855.31 Gt –
(µ,λ) mode, i.e., parents and offspring do not coexist. Self- load shedding 1463790.80
adaptation of search control parameters have been tried. The
Table 2: Optimized thermal commitment.
population size, the standard deviation of the mutation
operator, and the tournament size have been automatically
adjusted [9]. The mutation perturbation mechanism is
implemented as follows: Três Marias
x' = x + ε , for ε ~ N (0, σ 2 ) (25) Run-of-
where Sobradinho the-river
x : decision variables vector prior to mutation; Reservoir
x’ : decision variables vector after mutation. Itaparica
International Journal of Innovations in Energy Systems and Power (Vol. 4 no. 2, October 2009) 24 of 34
Data (I) (II) (III) (IV) (V) (VI) (VII) in the feasible interval of the decision variable. Elitism has
Max
Vol. 19528 34116 10782 900 260 128 3944 always saved the best two individuals from the previous
hm3
Min
generation.
Vol.
4250 5447 7238 900 260 128 3944
Therefore, six combinations of selection and crossover
a0 5.3037 3.741790 2.75813 2.515 2.303 2.515 1.38
×102 ×102 ×102 ×102 ×102 ×102 ×102
procedures have been tested 30 times each. Results are
a1 4.3359 1.39669 6.76489
--- --- --- --- presented in Table 6, in which the symbols RU, RP, RI, TU,
×10−3 ×10−3 ×10−3
a2 −2.4529 −5.35159 −8.86837 TP, and TI stand for the specific selection and crossover
--- --- --- ---
a3
×10−7
8.8877
×10−8
1.15599
×10−7
7.06791
operators.
--- --- --- ---
×10−12 ×10−12 ×10−11
a4 −1.3347 −9.54599 −2.23985
×10−16 ×10−18 ×10−15
--- --- --- --- RU RP RI TU TP TI
a0* 5.10037 3.606096 2.515 2.303 1.3412 1.29044 1.3721 2.05 2.25 1.08 1.88 2.03 1.90
×102 ×102 ×102 ×102 ×102 ×102 ×102 Mean
×108 ×108 ×1016 ×108 ×108 ×108
a1* 1.92841 1.24821 3.31878 2.07974 2.47288
×10−3 ×10−3
--- ---
×10−3 ×10−3 ×10−3 1.08 2.85 8.99 1.07 3.79 5.99
Deviation
a2* −1.74094 −1.27803 −3.09259 −5.27068 −3.22059 ×107 ×107 ×1015 ×107 ×107 ×106
--- ---
×10−7 ×10−7 ×10−7 ×10−8 ×10−7 1.91 1.95 4.91 1.71 1.71 1.80
a3* 1.2127 9.302374 2.15278 6.66456 2.28884 Min.
×10−11 ×10−12
--- ---
×10−11 ×10−13 ×10−11
×108 ×108 ×1014 ×108 ×108 ×108
a4* −3.24195
×10−16
−2.63114
×10−16
--- ---
−5.9295
×10−16
−2.23117
×10−17
−5.81037
×10−17
Table 6: Results from GA.
Produc-
0.008564 0.009025 0.008927 0.00906 0.0088 0.009035 0.009025
tivity
max
As usual, proportional selection (R) has led to premature
turb.
924 4278 3306 2200 2144 2400 2796 convergence. The minimum cost solution (171 millions) has
flow
m3/s been found by tournament selection and uniform crossover.
Max
outflow
1386 6417 4959 3300 3216 3600 4194 Figure 2 presents the corresponding stored volume of the
Min
500 640 640 640 640 640 650
reservoirs along the 24 months.
outflow 100%
Table 4: Hydraulic features of hydro plants (ai and ai* stand 90%
60%
Storage
50%
Plant (I) (II) (III) (IV) (V) (VI) (VII)
Month 40%
International Journal of Innovations in Energy Systems and Power (Vol. 4 no. 2, October 2009) 25 of 34
Table 8: Results from SA.
Figure 3 shows the least cost operation policy with PS, which
has been obtained using combination C5 for the control Figure 4 shows the best result from the application of SA,
parameters. This minimum cost is equal to R$ 134.33 which has been obtained with combination C6 of search
millions. parameters. Minimum cost in this case is equal to R$ 128.29
100%
millions.
90%
80%
100%
70%
90%
60%
80%
Storage
50%
70%
40%
60%
Storage
30%
50%
20%
40%
10%
30%
0%
May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr 20%
C1 C2 C3 C4 C5 C6
7.00 2.07 4.74 1.62 2.07 1.46
Mean ×108 ×108 ×108 ×108 ×108 ×108
1.02 5.60 1.14 1.56 5.62 1.05
Deviation ×108 ×107 ×108 ×107 ×107 ×107
5.36 1.45 2.92 1.36 1.46 1.28 Fig. 5: Percentage of the reservoirs’ storage (C2).
Min. ×108 ×108 ×108 ×108 ×108 ×108
International Journal of Innovations in Energy Systems and Power (Vol. 4 no. 2, October 2009) 26 of 34
E. Comparisons 100%
90%
of the seach space, the turbined flows can be discretized with 70%
Storage
discretized search space would be equal to 50%
30%
have been visited by GA, PS, SA, and ES are 288,000, 20%
An overall comparison among the four metaheuristics is Fig. 6: Percentage of the reservoirs’ storage (GRG).
summarized in Table 10. Although depending on an
appropriate starting point, the superior performance of Besides, setting the GRG control parameters has not been
simulated annealing with respect to minimum and average easier. Similar solutions have been achieved by GRG without
costs is noticeable. Regarding robustness, which can be the need for the interactive optimization process when GRG is
noticed from the variability coefficient (deviation over mean), initialized with solutions provided by the metaheuristics.
GA has been the most efficient. Moreover, results from ES Because GRG needs a good starting point, this is a very
show that self-tuning of search parameters has not produced efficient way to hybridize these optimization techniques, using
competitive results for this particular application. All test the ability of metaheuristics to explore different regions of the
results in Table 10 have reached feasibility. search space and the local search capacity of GRG.
When comparing the solutions provided by GRG and by
GA-TU PS-C5 SA-C6 ES-C2 the metaheuristics, it can be verified that all methods agree
Mean regarding the operation policy for the reservoirs of Três
1.8779 1.7624 1.4571 8.5995
(×108) Marias and Sobradinho. Notice that the solutions for these
hydro plants follow the inflows seasonality, with little
std.dev.
1.0701 2.3419 1.0538 78.068 variation on the turbined flows. On the other hand, there is no
(×107)
general agreement with respect to the operation of Itaparica.
Min (×108) 1.7140 1.3433 1.2829 2.6501 The GRG method has preserved its stored volume close to
100% most of the time (except for the last intervals, which are
dev/mean 5.70% 13.29% 7.23% 90.8% not important because of the end of the optimization horizon).
Because Itaparica has a small reservoir compared with the
Table 10: Global comparison.
other two hydro plants, it has been harder for the
metaheuristics to coordinate its operation policy. That has
Even PS, although not having achieved the best
been the reason why GRG has slightly improved the solutions
performances with respect to the minimum solution, average
cost, or robustness, has been the easiest algorithm to adjust. provided by the metaheuristics.
Regarding this particular aspect, the genetic algorithm has
been the hardest to set.
As a benchmark, the Generalized Reduced Gradient (GRG) V. CONCLUSIONS
method implemented in the Frontline Solver Platform This work has compared different metaheuristics when
(www.solver.com) has achieved a minimum cost of R$ 116.68 applied to the long-term hydrothermal coordination problem.
millions after the same initialization adopted for simulated The selected metaheuristics have been chosen based on
annealing. However, this result has been obtained by a manual previous experience of their application to large-scale
interactive process in which the decision variables related to optimization problems. The results presented in this paper
each reservoir are initially optimized considering one reservoir have confirmed their convergence robustness and the
at a time (with the other variables frozen). Afterwards, with solutions quality compared with a classical optimization
the values from the previous solutions, the decision variables technique.
are optimized considering pairs of reservoirs. Finally, the The tuning of search parameters has been harder for the
variables are optimized altogether (Fig. 6). That has been the Generalized Reduced Gradient method. A practical procedure
only way to allow GRG to converge to meaningful solutions, for incorporating hydrology, load, and fuel costs uncertainties
which is not practical for a larger number of reservoirs. based on multi-scenario optimization is an ongoing project.
VI. ACKNOWLEDGMENT
This work was partially supported by the Brazilian
Research Council (CNPq).
International Journal of Innovations in Energy Systems and Power (Vol. 4 no. 2, October 2009) 27 of 34
VII. REFERENCES
International Journal of Innovations in Energy Systems and Power (Vol. 4 no. 2, October 2009) 28 of 34