Professional Documents
Culture Documents
Research Article
Steady Modeling for an Ammonia Synthesis Reactor Based on
a Novel CDEAS-LS-SVM Model
Copyright 2014 Zhuoqian Liu et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
A steady-state mathematical model is built in order to represent plant behavior under stationary operating conditions. A novel
modeling using LS-SVR based on Cultural Differential Evolution with Ant Search is proposed. LS-SVM is adopted to establish the
model of the net value of ammonia. The modeling method has fast convergence speed and good global adaptability for identification
of the ammonia synthesis process. The LS-SVR model was established using the above-mentioned method. Simulation results verify
the validity of the method.
specific internal mechanism because a lot of parameters are been realized by LuHua Inc., a medium fertilizers factory of
unknown in real industrial process. YanKuang Group, China.
In order to achieve the required accuracy of the model, Figure 1 represents a flow sheet for the ammonia syn-
some researches focus on the novel modeling methods thesis process. The ammonia synthesis reactor is a one-axial
combining some heuristic methods such as ANN (Artificial flow and two-radial flow three-bed quench-type unit [11].
Neural Network), LS-SVM (Least Squares Support Vector Hydrogen-nitrogen mixture is reacted in the catalyst bed
Machine) with Evolutionary Algorithm, for example, genetic under high temperature and pressure. The temperature in
algorithm, ant colony optimization (ACO), particle swarm the reactor is sustained by the heat of reaction because the
optimization (PSO), differential evolution (DE), and so forth. reaction is exothermic [1]. The reaction of ammonia synthesis
DE is one of the most popular algorithms for this problem and process contains
has been applied in many fields. Sacco and Hendersonb [5]
3 1
introduced a variant of the differential evolution algorithm H + N NH3 + Q. (1)
with a new mutation operator based on a topographical 2 2 2 2
heuristic, and used it to solve the nuclear reactor core design The reaction is limited by the unfavorable position of
optimization problem. Rout et al. [6] proposed a simple but the chemical equilibrium and by the low activity of the
promising hybrid prediction model by suitably combining promoted iron catalysts with high pressure and temperature
an adaptive autoregressive moving average architecture and [12]. In general, no more than 20% of the synthesis gas is
differential evolution for forecasting of exchange rates. Ozcan converted into ammonia per pass even at high pressure of
et al. [7] carried out the cost optimization of an air cooling 30 MPa [12]. As the ammonia reaction is exothermic, it is
system by using Lagrange multipliers method, differential necessary for removing the heat generated in the catalyst bed
evolution algorithm and particle swarm optimization for by the progress of the reaction to obtain a reasonable overall
various temperatures and mass flow rates. The results showed conversion rate as same as to protect the life of the catalyst
that the method gives high accuracy results within a short [13]. The mixture gas from the condenser is divided into two
time interval. Zhang et al. [8] proposed a hybrid differential parts Q1 and Q2 to go to the converter. The first cold shot
evolution algorithm for the job shop scheduling problem with Q1 is recirculated to the annular space between the outer
random processing times under the objective of minimizing shell reactor and catalyst bed from the top to the bottom
the expected total tardiness. Arya and Choube [9] described to refrigerate the shell and remove the heat released by the
a methodology for allocating repair time and failure rates to reaction. Then the gas Q1 from the bottom of reactor goes
segments of a meshed distribution system using differential through the preheater and is heated by the counter-current
evolution technique. Xu et al. [10] proposed a model of flowing reacted gas from waste heat boiler. Q1 gas is divided
ammonia conversion rate by LS-SVM and a hybrid algorithm into 4 cold quench gas (q1, q2, q3, and q4) and Q2 gas for
of PSO and DE is described to identify the hyper-parameters mixing with the gas between consecutive catalyst beds to
of LS-SVM. quench the hot spots before entry to the subsequent catalyst
To describe the relationship between net value of ammo- beds. The hot spot temperatures (TIRA705, TIRA712N, and
nia in ammonia synthesis reactor and the key operational TIRA714) represent the highest reaction temperatures at each
parameters, least squares support vector machine is employed stage of the catalyst bed.
to build the structure of the relationship model, in which a Figure 2 represents the ammonia synthesis unit. The
novel algorithm called CDEAS is proposed to identify the reacted gas including N2 , H2 , NH3 , and inert gas after reactor
parameters. The experiment results showed that the proposed passes through the waste heat boiler. Then it goes through
CDEAS-LS-SVM optimizing model is very effective of being the preheater and the water cooler to be further cooled. Part
used to obtain the optimal operational parameters of ammo- of the ammonia is condensed and separated by ammonia
nia synthesis converter. separator I. Inert gas from the ammonia synthesis loop are
The remaining of the paper is organized as follows. ejected by purge gas from separator to prevent accumulation
Section 2 describes the ammonia synthesis production pro- of inert gas in the system. The fresh feed gas is produced
cess. Section 3 proposes a novel Cultural Differential Evolu- by the Texaco coal gasification air separation section, a
tion with Ant Colony Search (CDEAS) algorithm. Section 4 process that converts the Coal Water Slurry into synthesis
constructs a model using LS-SVM based on the proposed gas for ammonia. The fresh gas consists of hydrogen and
CDEAS algorithm. Section 5 presents the experiments and nitrogen in stoichiometric proportions of 3 : 1 approximately
computational results and discussion. Finally, Section 6 sum- and mixes with small amounts of argon and methane. The
marizes the above results and presents several problems fresh gas which passes compressor is compounded with the
which remain to be solved. recycle gas which comes from the circulator, and then the
mixture goes through oil separator and condenser. Mixture
gas is further cooled by liquid ammonia and goes through
2. Ammonia Synthesis Production Process ammonia separator II to separate the partial liquid ammonia,
and then it goes out with very few ammonia. The liquid
A normal ammonia production flow chart includes the ammonia from ammonia separator I and separator II flows
synthesis gas production, purification, gas compression, and to the liquid ammonia jar. Mixture is heated in ammonia
ammonia synthesis. Ammonia synthesis loop is one of the condenser to about 25 C and flows to the reactor and the
most critical units in the entire process. The system has whole cycle starts again.
Mathematical Problems in Engineering 3
AR AR FIR PI
701 701-4 705 703
Condenser
Ammonia
TIRA705 reactor
Evaporator Ammonia
TIR712N
TI cooler
TIRA714
725
Ammonia
Synthesis separator II
gas
Compressor
Circulator
Waste heat
boiler Preheater Water Ammonia Oil
cooler separator I separator
Liquid
ammonia
(2) Mutation. DE employs the mutation operation to produce As an effective and powerful random optimization
a mutant vector called target vector corresponding to each method, DE has been successfully used to solve real world
individual after initialization. In iteration , the mutant problems in diverse fields both unconstrained and con-
vector of individual can be generated according to strained optimization problems.
certain mutation strategies. Equations (3)(7) indicate the
most frequent mutation strategies version, respectively:
3.2. Cultural Differential Evolution with Ant Search. As we
DE/rand/1
=
1 +
(2
3 ), (3) mentioned in Section 3.1, mutation factor , mutation strate-
, , ,
gies, and crossover factor have great influence on the bal-
= 1 ance of DEs exploration and exploitation ability. decides the
DE/rand/2
, + (2 , 3 , )
(4) amplification of differential variation; is used to control
+ (4 the possibility of the crossover operation; mutation strategies
, 5 , ) , have great influence on the results of mutation operation. In
= best,
+ (1 some literatures , , and mutation strategies are defined in
DE/best/1
, 2 , ) , (5)
advance or varied by some specific regulations. But the factors
DE/best/2
= best,
+ (1 , , and strategies are very difficult to choose since the prior
, 2 , ) knowledge is absent. Therefore, Ant Colony Search is used
(6)
to search the suitable combination of , , and mutation
+ (3
, 4 , ) , strategies adaptively to accelerate the global search. Some
researchers have found an inevitable relationship between
DE/rand-to-best/1 = , + (best, , )
the parameters (, , and mutation strategies) and the
(7)
optimization results of DE [1618]. However, the approaches
+ (1
, 2 , ) , above are not applying the most suitable , , and mutation
strategies simultaneously.
where 1 , 2 , 3 , 4 , and 5 are mutually exclusive integers
In this paper, based on the theory of Cultural Algorithm
randomly generated within the range [1, NP] which should
and Ant Colony Optimization (ACO), an improved Cul-
not be . is the mutation factor for scaling the difference
tural Differential Algorithm incorporation with Ant Colony
vector, usually bounded in [0, 2]. best is the best individual
Search is presented. In order to accelerate searching out the
with the best fitness value at generation in the population.
global solution, the Ant Colony Search is used to search
(3) Crossover. The individual and mutant vector are the optimal combination of and in subpopulation 1 as
well as mutation strategy in subpopulation 2. The framework
hybridized to compose the trial vector after mutation
of Cultural Differential Evolution with Ant Search is briefly
operation. The binomial crossover is adopted by the DE in
described in Figure 3.
the paper, which is defined as
if rand or = rand 3.2.1. Population Space. The population space is divided into
= { (8)
, otherwise, two parts: subpopulation 1 and subpopulation 2. The two
subpopulations contain equal number of the individuals.
where rand is a random number between in 0 and 1 dis- In subpopulation 1, the individual is set as ant at each gen-
tributed uniformly. The crossover factor is a probability eration. and are defined to be the values between [0, 1],
rate within the range 0 and 1, which influences the tradeoff {0.1}, = 1, 2, . . . , 10 and {0.1}, = 1, 2, . . . , 10.
between the ability of exploration and exploitation. rand is an Each of the ants chooses a combination of and according
integer chosen randomly in [1, ]. To ensure that the trial to the information which is calculated by the fitness function
vector ( ) differs from its corresponding individual ( ) by of ants. During search process, the information gathered by
at least one dimension, = rand is recommended. the ants is preserved in the pheromone trails . By exchanging
information according to pheromone, the ants cooperate with
(4) Selection. When a newly generated trial vector exceeds each other to choose appropriate combination of and .
its corresponding upper and lower bounds, it is reinitialized Then ant colony renews the pheromone trails of all ants.
within the presetting range uniformly and randomly. Then Then, the pheromone trail is updated in the following
the trial individual is compared with the individual , and equation:
the one with better fitness is selected as the new individual in
the next iteration:
subpopulation1
+1
if (
) (, ) ( + 1) = (1 1 ) () +
() ,
= { (9)
, otherwise. =1
(10)
(5) Termination. All above three evolutionary operations
continue until termination criterion is achieved, such as
the evolution reaching the maximum/minimum of function where 0 1 < 1 means the pheromone trail evaporation
evaluations. rate, = 1, 2, . . . , 10, = 1, 2; 1st parameter represents and
Mathematical Problems in Engineering 5
Belief space
(situational knowledge and
normative knowledge)
Acceptance Influence
function function
Knowledge
Subpopulation 1 exchange Subpopulation 2
Ant Search of F and CR Ant Search of mutation strategy
F CR
0.1 0.1
0.2 0.2
0.3 0.3
.. ..
. .
1.0 1.0
2nd parameter represents ; () is the quantity of the considerably, the probability of each ant chooses th value of
pheromone trail of ant , th parameter ( and ) in Figure 4 is set by
{ ()
() { if rand1 <
() = { () (12)
{
{ 1 if and fitness ( ) < fitness (best ) ,
{
{ {rand2 otherwise.
{0.5 if and fitness (best ) < fitness ( )
={ Figure 4 illustrates the relationship between pheromone
{
{
{ and fitness ( ) < fitness ( ) , matrix and ant path of and , where is a constant
{0 otherwise, which is defined as selection parameter and rand1 and rand2
(11) are two random values which are uniformly distributed in
[0, 1]. Selection of the values of and depends on the
where is the ant group that chooses th value as the pheromone of each path. According to the performance of
selection of th parameter; best denotes the best individual all the individuals, the individual is chosen by the most
of ant colony till th generation. appropriate combination of and in each generation.
In order to prevent the ants from being limited to one In subpopulation 2, the individual is set as ant at each
ant path and improve the possibility of choosing other paths generation. Mutation strategies which are listed at (3)(7) are
6 Mathematical Problems in Engineering
Mutation strategy
0.2
0.4
0.6
..
.
1.0
1 2 3 4 5
defined to be of the values {0.2, 0.4, 0.6, 0.8, 1.0}, respectively. considerably, the probability of each ant choosing th value
For example, 0.2 means the first mutation strategy equation of th parameter (mutation strategies) is set by
(3) is selected. Each of the ants chooses a mutation strategy
according to the information which is calculated by the fitness
function of ants. During search process, the information ()
gathered by the ants is preserved in the pheromone trails . {
{ if rand3 <
By exchanging information according to pheromone, the ants () = { () (15)
{
cooperate with each other to choose appropriate mutation {rand4 otherwise,
strategy. Then ant colony renews the pheromone trails of all
ants.
Then, the pheromone trail is updated in the following
equation: where is a constant which is defined as selection parameter
and rand3 and rand4 are two random values which are
uniformly distributed in [0, 1]. Selection of the values of
subpopulation2 mutation strategies depends on the pheromone of each path.
( + 1) = (1 2 ) () + () , (13) According to the performance of all the individuals, the
=1 individual is chosen by the most appropriate combination of
mutation strategies in each generation.
Figure 5 illustrates the relationship between pheromone
where 0 2 < 1 means the pheromone trail evaporation
matrix and ant path of mutation strategies.
rate and () is the quantity of the pheromone trail of ant ,
and are set by Step 3. To find the proper , , and mutation strategy, the
Ant Colony Search strategy is used in subpopulation 1 and
, if ,min < or (,min ) <
= { ,min subpopulation 2, respectively.
otherwise,
(16)
Step 4. According to acceptance function, choose good indi-
, if ,max > or (,max ) >
= { ,max viduals from subpopulation 1 and subpopulation 2, and then
otherwise. update the normative knowledge and situational knowledge.
3.2.3. Acceptance Function. Acceptance function controls the Step 5. Adopt the normative knowledge and situational
amount of good individuals which impact on the update of knowledge to influence each individual in population space
belief space [19]. In this paper, 30% of the individuals in through the influence functions, and generate two corre-
the belief space are replaced by the good ones in population sponding subpopulations.
space.
Step 6. Select individuals from subpopulation 1 and subpop-
3.2.4. Influence Function. In the CDEAS, situational knowl-
ulation 2, and update the belief spaces including the two
edge and normative knowledge are involved to influence each
knowledge sources for the next generation.
individual in the population space, and then population space
is updated.
The individuals in population space are updated in the Step 7. If the algorithm reaches the given times, exchange
following equation: the knowledge of , , and mutation strategy between
subpopulation 1 and subpopulation 2; otherwise, go to Step 8.
{
, + (0.5, 0.3) (2
3
)
,
rand,
{
{
,
{
{
{
{ if , , , , Step 8. If the stop criteria are achieved, terminate the itera-
{
{
{
{
tion; otherwise, go back to Step 2.
{
{, (0.5, 0.3) (2 , 3 , ) rand,
{
{
{
{
{
{ if ,
> , ,
< , 3.3. Simulation Results of CDEAS. The proposed CDEAS
+1
, ={
{
{
+ (0.5, 0.3) ( 3 rand, algorithm is compared with original DE algorithm. To get
{ 1 )
{
{
, , the average performance of the CDEAS algorithm 30 runs
{
{
{
{ if ,
, ,
, on each problem instance were performed and the solution
{
{
{
{ quality was averaged. The parameters of CDEAS and original
{
{
{ 1 , (0.5, 0.3) ( 3
)
,
rand,
{
{
DE algorithm are set as follows: the maximum evolution
{ if , > , , > , generation is 2000; the size of the population is 50; for original
DE algorithm = 0.3 and = 0.5; for CDEAS, the size
1 + ( 1
) rand, if , > of both two subpopulations is 25; the initial and are
{
{
, ,
+1 {
randomly selected in (0, 1) and the initial mutation strategy
, = {1 , (1 , ) rand, if , < is DE/rand/1; the interval information exchanges between the
{
{
two subpopulations is 50 generations; the thresholds =
{1 , + ( ) rand, if < , < ,
= 0.5 and 1 = 2 = 0.1.
(17)
To illustrate the effectiveness and performance of CDEAS
where is a constant of 0.2. algorithm for optimization problems, a set of 18 representa-
tive benchmark functions which were listed in the appendix
3.2.5. Knowledge Exchange. After steps, the and were employed to evaluate them in comparison with original
of subpopulation 2 are replaced by the suitable and DE. The test problems are heterogeneous, nonlinear, and
calculated by subpopulation 1 and the mutation strategy numerical benchmark functions and the global optimum for
of subpopulation 1 is displaced by the suitable mutation 2 , 4 , 7 , 9 , 11 , 13 , and 15 is shifted. Functions 1 7 are
strategy calculated by subpopulation 2 simultaneously. So unimodal and functions 8 18 are multimodal. The detailed
the and and mutation strategy are varying in the two principle of functions is presented in [11]. The comparisons
subpopulations to enable the individuals to converge globally results of CDEAS and original DE algorithm are shown in
and fast. Table 4 of the appendix. The experimental results of original
DE and CDEAS algorithm on each function are listed in
3.2.6. Procedure of CDEAS. The procedure of CDEAS is Table 1. Mean, best, worst, std., success rate, time represent
proposed as follows. the mean minimum, best minimum, worst minimum, the
standard deviation of minimum, the success rate, and the
Step 1. Initialize the population spaces and the belief spaces; average computing time in 30 trials, respectively.
the population space is divided into subpopulation 1 and From simulation results of Table 1 we can obtain that
subpopulation 2. CDEAS reached the global optimum of 2 and 7 in all trials,
and the success rate reached 100% of functions 1 , 2 , 3 , 4 ,
Step 2. Evaluate each individuals fitness. 6 , 7 , and 18 . For most of the test functions, the success
8 Mathematical Problems in Engineering
Table 1: The comparison results of the CDEAS algorithm and original DE algorithm.
Original DE CDEAS
Sphere function 1
Best 1.1746 1065 5.0147 1079
Worst 1.0815 1023 9.3244 1075
Mean 3.6052 1025 1.6390 1075
Std. 1.9746 1024 2.2315 1075
Success rate (%) 100 100
Times (s) 1.8803 14.6017
Shifted sphere function 2
Best 0 0
Worst 8.0779 1028 0
Mean 3.3658 1029 0
Std. 1.5078 1028 0
Success rate (%) 100 100
Times (s) 2.1788 18.1117
Schwefels Problem 1.2 3
Best 2.4386 1065 3.0368 1078
Worst 2.4820 1022 9.2902 1073
Mean 8.2736 1024 7.2341 1074
Std. 4.5316 1023 2.0187 1073
Success rate (%) 100 100
Times (s) 3.1647 24.1178
Shifted Schwefels Problem 1.2 4
Best 0 0
Worst 5.6545 1027 3.4331 1027
Mean 2.0868 1028 1.8848 1028
Std. 1.0323 1027 7.9813 1028
Success rate (%) 100 100
Times (s) 3.3956 27.7058
Rosenbrocks function 5
Best 13.0060 5.2659
Worst 166.1159 139.1358
Mean 70.9399 39.4936
Std. 40.0052 31.2897
Success rate (%) 86.67 96.67
Times (s) 1.9594 16.7233
Schwefels Problem 1.2 with noise in fitness 6
Best 3.1344 1039 3.98838 1049
Worst 3.61389 1036 1.6124 1043
Mean 5.7744 1037 7.4656 1045
Std. 9.5348 1037 2.9722 1044
Success rate (%) 100 100
Times (s) 3.2141 24.2426
Shifted Schwefels Problem 1.2 with noise in fitness 7
Best 0 0
Worst 0 0
Mean 0 0
Mathematical Problems in Engineering 9
Table 1: Continued.
Original DE CDEAS
Std. 0 0
Success rate (%) 100 100
Times (s) 3.3374 28.5638
Ackleys function 8
Best 7.1054 1015 3.5527 1015
Worst 4.8999 107 1.3404
Mean 1.6332 108 0.1763
Std. 8.9457 108 0.4068
Success rate (%) 100 83.33
Times (s) 2.4820 20.9353
Shifted Ackleys function 9
Best 7.1054 1015 3.5527 1015
Worst 0.9313 0.9313
Mean 0.0310 0.0620
Std. 0.1700 0.2362
Success rate (%) 96.67 93.33
Times (s) 2.7337 21.6841
Griewanks function 10
Best 0 0
Worst 0.0367 0.0270
Mean 0.0020 0.0054
Std. 0.0074 0.0076
Success rate (%) 90 56.67
Times (s) 2.535 20.7793
Shifted Griewanks function 11
Best 0 0
Worst 0.0319 0.0343
Mean 0.0056 0.0060
Std 0.0089 0.0088
Success rate (%) 80 76.67
Times (s) 2.7768 22.8541
Rastrigins function 12
Best 8.1540 1.9899
Worst 35.5878 12.9344
Mean 20.3594 6.5003
Std. 6.3072 2.6612
Success rate (%) 3.33 90
Times (s) 2.7264 22.3237
Shifted Rastrigins function 13
Best 5.9725 0.9949
Worst 36.9923 6.7657
Mean 19.4719 8.2581
Std. 8.9164 3.8680
Success rate (%) 16.67 76.67
Times (s) 2.9313 23.8838
Noncontiguous Rastrigins function 14
Best 20.7617 3.9949
Worst 29.9112 11.9899
Mean 25.4556 8.1947
10 Mathematical Problems in Engineering
Table 1: Continued.
Original DE CDEAS
Std. 2.9078 2.2473
Success rate (%) 0 86.67
Times (s) 3.1663 25.5374
Shifted noncontiguous Rastrigins function 15
Best 0 0
Worst 16 6
Mean 6.7666 1.5333
Std. 3.4509 1.8519
Success rate (%) 40 96.67
Times (s) 3.3374 25.9430
Schwefels function 16
Best 118.4387 236.8770
Worst 710.6303 1362.0521
Mean 357.61725 676.4166
Std. 144.41244 324.2317
Success rate (%) 90 40
Times (s) 2.5028 19.0009
Schwefels Problem 2.21 17
Best 0.1640 0.3254
Worst 4.5102 4.7086
Mean 1.1077 1.9849
Std. 0.8652 1.16418
Success rate (%) 53.33 23.33
Times (s) 2.3806 19.2505
Schwefels Problem 2.22 18
Best 1.2706 1035 8.5946 1045
Worst 1.6842 1034 1.8362 1042
Mean 6.1883 1035 2.6992 1043
Std. 3.4937 1035 4.6257 1043
Success rate (%) 100 100
Times (s) 2.6297 20.8573
rate of CDEAS is higher in comparison with original DE. 4. Model of Net Value of Ammonia Using
Moreover, CDEAS gets very close to the global optimum in CDEAS-LS-SVM
some other functions 1 , 3 , 4 , 6 , and 18 . It also presents
that the mean minimum, best minimum, worst minimum, 4.1. Auxiliary Variables Selection of the Model. There are some
the standard deviation of minimum, and the success rate of process variables which have the greatest influence on the net
CDEAS algorithm are clearly better than the original DE for value of ammina, such as system pressure, recycle gas flow
functions 1 , 3 , 4 , 5 , 6 , 12 , 13 , 14 , 15 , and 18 although rate, feed composition (H/N ratio), ammonia and inert gas
the computing time of CDEAS is longer than that of original cencetration in the gas of reactor inlet, hot spot temperatures,
DE because of its complexity. and so forth. The relations between the process variables
The convergence figures of CDEAS comparing with are coupling and the operational variables interact with each
original DE for 18 instances are listed as Figure 6. other.
From Figure 6 one can observe that the convergence The inlet ammonia concentration is an important process
speed of CDEAS is faster than original DE for 1 , 2 , 3 , 4 , variable which is beneficial to operation-optimization but the
6 , 7 , 11 , 12 , 13 , 14 , 15 , and 18 . device of online catharometer is very expensive. According
All these comparisons of CDEAS with original DE algo- to the mechanism and soft sensor model, a IIO-BP model
rithm have shown that CDEAS is a competitive algorithm was built to get the more accurate value of the inlet ammonia
to solve all the unimodal function problems and most of concentration [20]
the multimodal function optimization problems listed above.
As shown in the descriptions and all the illustrations before,
CDEAS is efficacious on those typical function optimizations. (NH3 ) = NH3 OUT NH3 IN . (18)
Mathematical Problems in Engineering 11
log(fitness value)
log(fitness value)
5
20
30 10
40 15
50
20
60
70 25
80 30
0 400 800 1200 1600 2000 0 400 800 1200 1600 2000
Evolution generation Evolution generation
log(fitness value)
0
log(fitness value)
20 5
30
10
40
15
50
60 20
70 25
80 30
0 400 800 1200 1600 2000 0 400 800 1200 1600 2000
Evolution generation Evolution generation
log(fitness value)
8 10
7 15
6 20
5 25
4 30
3 35
2 40
1 45
0 400 800 1200 1600 2000 0 400 800 1200 1600 2000
Evolution generation Evolution generation
5 4
10 6
15 8
20 10
25 12
30 14
35 16
0 400 800 1200 1600 2000 0 400 800 1200 1600 2000
Evolution generation Evolution generation
Original DE Original DE
CDEAS CDEAS
Figure 6: Continued.
12 Mathematical Problems in Engineering
log(fitness value)
2
4
4
6
6
8 8
10 10
12 12
14 14
16 16
0 400 800 1200 1600 2000 0 400 800 1200 1600 2000
Evolution generation Evolution generation
log(fitness value)
1 2.2
2
0 1.8
1.6
1 1.4
2 1.2
1
3 0.8
0 400 800 1200 1600 2000 0 400 800 1200 1600 2000
Evolution generation Evolution generation
log(fttness value)
2.2 2.2
2 2
1.8 1.8
1.6 1.6
1.4 1.4
1.2 1.2
1 1
0.8 0.8
0 400 800 1200 1600 2000 0 400 800 1200 1600 2000
Evolution generation Evolution generation
log(fttness value)
2 3.6
3.4
1.5 3.2
3
1 2.8
0.5 2.6
2.4
0 2.2
0 400 800 1200 1600 2000 0 400 800 1200 1600 2000
Evolution generation Evolution generation
Original DE Original DE
CDEAS CDEAS
Figure 6: Continued.
Mathematical Problems in Engineering 13
0 1.5
log(fitness value)
log(fitness value)
10
1
20
0.5
30
0
40
50 0.5
0 400 800 1200 1600 2000 0 400 800 1200 1600 2000
Evolution generation Evolution generation
Original DE Original DE
CDEAS CDEAS
From the analysis discussed above, some important vari- where is a [ 1] vector of ones, is the transpose of a
ables have significant effects on the net value of ammonia. matrix or vector, is a weight vector, 0 means the model
By discussion with experienced engineers and taking into offset, and is regression vector.
consideration a priori knowledge about the process, the is Mercer kernel matrix, which is defined as
system pressure, recycle gas flow rate, the H/N ratio, hot-spot
temperatures in the catalyst bed, and ammonia and methane 1,1 1,
concentration in the recycle gas are identified as the key = ( ... d .. ) ,
. (20)
auxiliary variables to model net value of ammonia which is ,1 ,
listed in Table 2.
where , is defined by kernel function.
4.2. Modeling the Net Value of Ammonia Using CDEAS- There are several kinds of kernel functions, such as
LS-SVM. LS-SVM is an alternate formulation of SVM, hyperbolic tangent, polynomial, and Gaussian radial basis
which is proposed by Suykens. The e-insensitive loss func- function (RBF) which are commonly used. Literatures have
tion is replaced by a squared loss function, which con- proved that RBF kernel function has strong generalization,
structs the Lagrange function by solving the problem linear so in this study RBF kernel was used:
Karush Kuhn Tucker (KKT) 2 2
, = | | /2 , (21)
RE: relative error; MAE: mean absolute error; MSE: mean square error.
As we can see from (19)(21), only two parameters respectively. The results indicate that the proposed CDEAS-
(, ) are needed for LS-SVM. It makes LS-SVM problem LS-SVM model has a good tracking precision performance
computationally easier than SVR problem. and guides production better.
Grid search is a commonly used method to select the
parameters of LS-SVM, but it is time-consuming and inef-
ficient. CDEAS algorithm has strong search capabilities, and 6. Conclusion
the algorithm is simple and easy to implement. Therefore, this
paper proposes the CDEAS algorithm to calculate the best In this paper, an optimizing model which describes the
parameters (, ) of LS-SVM. relationship between net value of ammonia and key opera-
tional parameters in ammonia synthesis has been proposed.
Some representative benchmark functions were employed to
5. Results and Discussion evaluate the performance of a novel algorithm CDEAS. The
obtained results show that CDEAS algorithm is efficacious
Operational parameters such as H2 , CH4 , and were for solving most of the optimization problems comparisons
collected and acquired from plant DCS from the year 2011- with original DE. Least squares support vector machine
2012. In addition, data on the inlet ammonia concentration is used to build the model while CDEAS algorithm is
of recycle gas NH3 were simulated by mechanism and soft employed to identify the parameters of LS-SVM. The sim-
sensor model [20]. ulation results indicated that CDEAS-LS-SVM is superior
The extreme values are eliminated from the data using the to other models (BP-NN, LS-SVM, and DE-LS-SVM) and
3 criterion. After the smoothing and normalization, each meets the requirements of ammonia synthesis process. The
data group is divided into 2 parts: 223 groups of training CDEAS-LS-SVM optimizing model makes it a promising
samples which are used to train model while 90 groups of candidate for obtaining the optimal operational parameters
testing samples which are valuing the generalization of the of ammonia synthesis process and meets the maximum
model for identifying the parameters of the LS-SVM, the benefit of ammonia synthesis production.
kernel width parameter, and the weight vector.
BP-NN, LS-SVM, and DE-LS-SVM are also used to
model the net value of ammonia, respectively. BP-NN is Appendix
a 13-15-1 three-layer network with back-propagation algo-
rithm. LS-SVM gains the (, ) with grid-search and cross- (1) Sphere function
validation. The parameter settings of CDEAS-LS-SVM are
the same as those in the benchmark tests. Each model is run
30 times and the best value is shown in Table 3. Descriptive 1 () = 2 ,
statistics of training results and testing results of model =1 (A.1)
include the relative error, absolute error, and mean square
error. The performance of the four models is compared as = [0, 0, . . . , 0] : the global optimum.
shown in Table 3. The training and testing results of four
models are illustrated in Figure 7. (2) Shifted sphere function
Despite the fact that the training error using BP-NN is
smaller than that using CDEAS-LS-SVM, which is because
BP-NN is overfitting to the training data, the mean square
0.116 0.115
0.115 0.114
0.116 0.115
0.115 0.114
Net value of ammonia (%)
0.116 0.115
0.115 0.114
Net value of ammonia (%)
0.114
0.113
0.113
0.112 0.112
0.111 0.111
0.11 0.11
0.109
0.109
0.108
0.107 0.108
0.106 0.107
0 50 100 150 200 250 0 10 20 30 40 50 60 70 80 90
Sample number Sample number
0.116 0.115
0.115 0.114
Net value of ammonia (%)
Net value of ammonia (%)
0.114
0.113
0.113
0.112 0.112
0.111 0.111
0.11 0.11
0.109
0.109
0.108
0.107 0.108
0.106 0.107
0 50 100 150 200 250 0 10 20 30 40 50 60 70 80 90
Sample number Sample number
Figure 7: The analyzed results, training results, and testing results of BP-NN, LS-SVM, DE-LS-SVM, and CDEAS-LS-SVM.
16 Mathematical Problems in Engineering
Table 4: Global optimum, search ranges, and initialization ranges of the test functions.
(3) Schwefels Problem 1.2 (7) Shifted Schwefels Problem 1.2 with noise in fitness
2 2
3 () = ( ) , 7 () = (( ) ) (1 + 0.4 | (0, 1)|) ,
=1 =1 (A.3) =1 =1
(A.7)
= [0, 0, . . . , 0] : the global optimum. = ,
= [0, 0, . . . , 0] : the global optimum. = [418.9829, 418.9829, . . . , 418.9829]: the global optimum.
(A.16)
(11) Shifted Griewanks function (17) Schwefels Problem 2.21
2
18 () = max { , 1 } ,
11 () = cos + 1, (A.17)
=1 4000 =1 = [0, 0, . . . , 0] : the global optimum.
(A.11)
= , (18) Schwefels Problem 2.22
= [1 , 2 , . . . , ] : the shifted global optimum.
17 () = + ,
=1 =1 (A.18)
(12) Rastrigins function
= [0, 0, . . . , 0] : the global optimum.
12 () = (2 10 cos (2 ) + 10) , Conflict of Interests
=1 (A.12)
= [0, 0, . . . , 0] : the global optimum. The authors declare that there is no conflict of interests
regarding the publication of this paper.
(13) Shifted Rastrigins function
Acknowledgments
13 () = (2 10 cos (2 ) + 10) , The authors are grateful to the anonymous reviewers for
=1 giving us helpful suggestions. This work is supported by
(A.13) National Natural Science Foundation of China (Grant nos.
= ,
61174040 and 61104178) and Fundamental Research Funds for
= [1 , 2 , . . . , ] : the shifted global optimum. the Central Universities, Shanghai Commission of Science
and Technology (Grant no. 12JC1403400).
(14) Noncontiguous Rastrigins function
References
14 () = (2 10 cos (2 ) + 10) , [1] M. T. Sadeghi and A. Kavianiboroujeni, The optimization
=1 of an ammonia synthesis reactor using genetic algorithm,
International Journal of Chemical Reactor Engineering, vol. 6, no.
1
{
{ < 1, article A113, 2009.
= { round (2 ) 2 for = 1, 2, . . . , , [2] S. S. E. H. Elnashaie, A. T. Mahfouz, and S. S. Elshishini,
{ 1
, Digital simulation of an industrial ammonia reactor, Chemical
{ 2 2 Engineering and Processing, vol. 23, no. 3, pp. 165177, 1988.
[3] M. N. Pedemera, D. O. Borio, and N. S. Schbib, Steady-State
= [0, 0, . . . , 0] : the global optimum.
analysis and optimization of a radial-flow ammonia synthesis
(A.14) reactor, Computers & Chemical Engineering, vol. 23, no. 1, pp.
S783S786, 1999.
(15) Shifted noncontiguous Rastrigins function [4] B. V. Babu and R. Angira, Optimal design of an auto-thermal
ammonia synthesis reactor, Computers & Chemical Engineer-
ing, vol. 29, no. 5, pp. 10411045, 2005.
15 () = (2 10 cos (2 ) + 10) , [5] W. F. Sacco and N. Hendersonb, Differential evolution with
=1
topographical mutation applied to nuclear reactor core design,
1 Progress in Nuclear Energy, vol. 70, pp. 140148, 2014.
{
{ < [6] M. Rout, B. Majhi, R. Majhi, and G. Panda, Forecasting of
= { round (2 ) 2 for = 1, 2, . . . , ,
{ 1 currency exchange rates using an adaptive ARMA model with
, differential evolution based training, Journal of King Saud
{ 2 2
University, vol. 26, no. 1, pp. 718, 2014.
= , [7] H. Ozcan, K. Ozdemir, and H. Ciloglu, Optimum cost of an
air cooling system by using differential evolution and particle
= [1 , 2 , . . . , ] : the shifted global optimum. swarm algorithms, Energy and Buildings, vol. 65, pp. 93100,
(A.15) 2013.
18 Mathematical Problems in Engineering
International
Journal of Journal of
Mathematics and
Mathematical
Discrete Mathematics
Sciences