Synthesis of Multicomponent Products Separation Sequences via

Oct 21, 2008 - This paper addresses the application of genetic programming (GP) to the synthesis of ... programming (GDP) model for the optimal synthe...
1 downloads 0 Views 129KB Size
Ind. Eng. Chem. Res. 2008, 47, 8815–8822

8815

Synthesis of Multicomponent Products Separation Sequences via Stochastic GP Method Xiao-Hong Wang* and Yu-Gang Li QingDao UniVersity of Science and Technology, QingDao, 266042, Shandong, China

This paper addresses the application of genetic programming (GP) to the synthesis of multicomponent products nonsharp distillation sequences, and the proposed method regards minimizing the annual total cost as an optimization objective and seeks the nonsharp separation optimal flow to give the major technological parameters of the important equipment which provide the accurate theoretical basis for actual production. In combination with the domain knowledge of chemical engineering, some evolutionary factors are improved, and a set of a special encoding method and solving strategy is proposed to deal with this kind of problem. The system structural variable is optimized by GP, and the continuous variable is optimized by the complex algorithm simultaneously. Because GP has the automatic searching function, the optimal solution can be automatically found including distillation, splitting, blending, and bypassing operations without any superstructures of nonsharp distillation sequences. Three illustrating examples are presented to demonstrate the effective computational strategies. 1. Introduction The synthesis of optimal distillation sequences with pure products has been studied very deeply over many years. Some distinct thermally coupled configurations were searched and for each of the distinct configurations, all the thermodynamically equivalent configurations can be easily drawn by Agrawal.1-3 Caballero and Grossmann4 proposed the generalized disjunctive programming (GDP) model for the optimal synthesis of thermally linked distillation columns. Grossmann5 gave a general comment about the nonlinear mixed-integer and disjunctive programming techniques. Ismai et al.6 proposed a systematic approach for the synthesis of combined separation/reaction systems which involves a phenomena-based representation formulated within a superstructure optimization framework, and a multifunctional mass/heat transfer-based process module is utilized as the building block of process alternatives. Grossmann et al.7 examine the impact of different representations and models, NLP, MINLP, and GDP, and discuss the synthesis of complex column configurations for zeotropic mixtures. Recently, Caballero and Grossmann8 improved a superstructure approach for synthesizing heat-integrated thermally coupled distillation sequences, which was based on a state-task approach instead of an equipment-based system. According to actual industrial production, the multicomponent products separation problem is very important. Nath9 and Muraki10 studied the synthesis of nonsharp distillation sequences with one feed and multiple multicomponent products by using the material allocation diagram (MAD). Liu et al.11 improved the individual material allocation diagram (IMAD) to solve this kind of problems with multifeed. Bamopoulos et al.12 proposed a two-stage approach based on the component recovery matrix (R-matrix) which is an algebraic extension of MAD, but only limited stream splitting is permitted in this study. Cheng and Liu13 devised some techniques for the systematic synthesis of initial sequences for nonsharp separations with heuristics. Liu and Xu14 improved the separation product matrix transforming method to solve the nonsharp distillation process with multifeed, then, the two-stage procedure for the synthesis of nonsharp * To whom correspondence should be addressed. E-mail: wxhlee@ sohu.com. Tel.: +86-532-84022026.

distillation flow was brought forward.15 Hu et al.16 gave an ordered heuristics method for the synthesis of nonsharp distillation sequences with simple, sharp distillation columns. Wehe and Westerberg17 described an algorithmic procedure for the synthesis of sharp distillation sequences with bypass. Aggarwal and Floudas18 addressed the process including heat integrated nonsharp columns with MINLP method. Recently, a stochastic optimization algorithm is used to solve separation synthesis problems. Zhang and Linninger19 gave a novel chromosome encompassing all column configurations to solve the mixed-integer nonlinear programming (MILNP) using a stochastic genetic algorithm (GA) for synthesizing multicomponent distillation networks. Leboreiro and Acevedo20 proposed an optimization framework for the synthesis and design of complex distillation sequences based on a modified GA coupled with a sequential process simulator. But they all did not solve the multicomponent products separation problem. In this paper, a new stochastic optimization algorithmsgenetic programming (GP)21 is proposed for the synthesis of multicomponent products nonsharp distillation sequences. GP can express the complex net structure of this kind of problems with its hierarchy and can determine the feasible solving area automatically without giving any superstructures. So, the global optimum can be searched automatically in the genetic evolutionary process. A set of solution strategy based on GP is suggested in this paper. GP is widely used in many fields as a kind of the stochastic optimization algorithms method. The principle and operations of GP are very similar to Genetic algorithm (GA), which starts its calculation from initialized populations and generates new populations using a variety of evolution operators to achieve the optimal solution according to the biological principles of natural selection. The difference between the two methods in nature is that GA uses the length-fixed character string to represent problems while GP adopts the dendriform code to express complex structures directly. So, the evolutionary operations of GP can also proceed on its dendriform code directly. To give an example that is illustrated in Figure 1, the following

10.1021/ie800610s CCC: $40.75  2008 American Chemical Society Published on Web 10/21/2008

8816 Ind. Eng. Chem. Res., Vol. 47, No. 22, 2008

Figure 1. Dendriform structure of expressions.

two mathematical expressions can be represented with particular dendriform structures of GP: (1) y ) a + b × x; (2) y ) a × exp(b × x). Because GP code is a kind of dendriform nonlinear data structures, it is a tree composed by some nodes and branches. The evolution operations, which include the operations of reproduction, crossover, and mutation for a single node or a whole branch of GP, can proceed in the dendriform structure of GP code directly. When GP code is used to represent the complex distillation sequence, the branches of the GP dendriform structure can be defined as streams and the nodes can be defined as unit operations (for example column, splitter, and blender). So, the combination of all nodes and branches for GP code can be used to express the actual nonsharp distillation structure directly. We can know from the aforementioned description that although GP is a domain-independent optimization algorithm, it needs domain knowledge to instruct the definition for its tree-like code. At the same time, evolution operations need to proceed in the relative professional domains to ensure the effectiveness and accuracy of the algorithm. So, combining with the professional knowledge of chemical engineering separation, the definition for the GP tree-like code and the evolutionary operations are specially restricted in this work in order to get an efficient optimization method for nonsharp distillation system synthesis. 2. Problem Statement The problem to be studied in this paper can be described as follows. A single feed stream of N-component mixture is given with the known conditions (i.e., composition, flow rate, temperature, and pressure). The problem is to synthesize an optimal separation flow that can separate the multicomponent mixture into the specified multicomponent product streams. The configurations can involve series and/or parallel arrangements of distillation columns, as well as stream splitting, blending, and bypassing. To be convenient for expression and calculation, the following assumptions are introduced in this work: (1) Each distillation column performs a “sharp” separation. (2) The maximum number of separators is N - 1. (3) Components in any stream are ranked according to their relative volatility order. To give an example, N components in a stream are arranged as E1, E2,..., EN, where E1 is the most volatile (lightest) component, E2 is the next lightest, and so forth down to the heaviest of all the components EN. (4) Heat integration among the distillation columns is not permitted. (5) The nonsharp separation flow is optimized in order to minimize annual total cost. 3. GP-Based Encoding Strategy 3.1. Nonsharp Distillation Sequence Encoding. To express the positions and the connection relationships of distillation,

splitting, blending and bypassing operations in GP tree-like code, three kinds of nodes (i.e., column node, splitter node, and blender node) are defined in this paper. Some relevant properties are set for these nodes to represent their features respectively, and the definition can be stated as follows: 3.1.1. Design for Column Node. Because the information of separation position (i.e., confirming for the light and heavy key component) is a very important property for column node, we use the corresponding ordinal number of light key component to represent it. For example, integer one denotes the light key component (i.e., no. 1 component) and the heavy key component is the adjacent one to the light key component for a sharp separation column. It requires N - 1 columns to separate the N-component mixture when only conventional columns with sharp separation are adopted. So, it corresponds to N - 1 different integer to represent cut property, and the relationship between the property and the corresponding separation mission is as follows. E1 | E2 | E3 · · · Ei|Ei+1 · · · EN-1 | EN 1 1

2 2

i i

N-1 N-1

(1)

Feed composition is another important information for column node. It is represented with two integers in this paper, one of them denotes the lightest component in the feed stream and the other indicates the heaviest component. Both are expressed by the corresponding component’s ordinal number. Figure 2a gives the definition for column node, where the feed composition parameters are expressed by symbol L and H, where L ) m and H ) n means that this feed stream contains components whose ordinal numbers range from m to n. The separation position information is expressed by “Separation ) k”, which means the light key component ordinal number is k, so, the heavy key component ordinal number is k + 1 for a sharp distillation column. Figure 2b describes the corresponding operation flow for this simple sharp distillation column that can separate this feed stream into two product streams. Distillate contains the components lighter than the light key component (i.e., from component m to k, including this light key component) and residue contains the components heavier than the heavy key component (i.e., from component k + 1 to n, including this light key component). It is specified that the information of stream cut position and feed composition are randomly selected for every column when the GP code is produced, so, the randomization of this algorithm can be assured. 3.1.2. Design for Splitter Node. A splitter can divide an inlet stream into several outlet streams that have the same compositions and physical properties but different flow rates. Similar to the definition for column node, it needs two kinds of information to express the properties for splitter node, too. First, two integers are used to describe the information of feed composition and the definition method for the two integers is the same as it for column node that is mentioned earlier. Other information needed for splitter node is the definition of the distribution coefficient for every outlet stream that denotes the ratio of an outlet streamflow rate to the inlet streamflow rate. To ensure the regularity of GP tree-like code, we limit that every splitter can only divide an inlet stream into two outlet streams in this paper. A real number is used to indicate the distribution coefficient of one outlet stream, at the same time, the distribution coefficient of another outlet stream is confirmed automatically, because the sum of the two distribution coefficients must be 1. Figure 3 gives the definition for splitter node where symbol “a” denotes the distribution coefficient of right side outlet stream and “1 - a” expresses the other outlet stream.

Ind. Eng. Chem. Res., Vol. 47, No. 22, 2008 8817

Figure 2. The sketch map of column node.

Figure 3. The sketch map of splitter node.

3.1.3. Design for Blender Node. Blender is very important in such nonsharp separation sequence. If the definition for it is random, that is to say, every blender is encoded randomly just as the same as that of the other two kinds of nodes, it will appear randomly at any position of GP code. Then, GP code becomes too complex and disordered. Moreover, it will result in the failure to express the mixed operation of any two or multiple feeds into the same blender. To express the special netlike structure of nonsharp separation procedure, a special encoding method for blender node is adopted in this paper, which sets all the blenders of GP code to be arranged in a kind of grid framework. The maximum layer number for the grid framework is usually confirmed according to the scale of every detailed problem. At the same time, the maximal number of blenders in

every layer for the blender grid framework is defined to the number of products. To denote the position of each blender in the framework, two integers are figured as the property of blender node. One of them denotes the layer number and the other indicates the ordinal number of the blender in this layer. The last layer of blender grid framework is named terminal products blender. To produce the prescribed mixture products, all streams of GP code will be collected to the several terminal products blenders, so, the number of terminal products blender must be the same as that of the products. 3.2. Generating GP Code. The whole GP code is based on the above-mentioned blender grid structure and every blender in it can be as a subroot node from which a subtree can be generated via growing downward randomly. So, many subtrees can be formed according to the same principle. When the whole GP code reaches the maximum depth limited by the algorithm, all the branches will be connected with several terminal blender nodes randomly to produce different mixture products and the whole algorithm tree ceases growth. That is to say, the whole GP tree-like code is finally formed based on the growth of subtree layer upon layer, which can be used to express the complex netlike structure of nonsharp distillation sequence.

Figure 4. Netlike structure of GP code: (S) splitter node; (C) column node; (M) blender node.

8818 Ind. Eng. Chem. Res., Vol. 47, No. 22, 2008

Figure 5. Crossover operation demonstration. Table 1. The Available Utilities utilities

pressure (Mpa)

price ($/t)

(1) cooling water (2) steam 1 (3). steam 2 (4) steam 3 (5) steam 4 (6) steam 5 (7) steam 6

0.1 0.35 0.7 1.0 1.7 4.0

0.0067 5.03 6.17 6.8 7.5 8.2 9.96

The detail process for generating GP code includes the following steps. First, a column node or a splitter node is selected randomly as GP code tree’s root node that expresses the original feed can be separated by a distillation column or be divided by a splitter. If a column node is selected as the root node, which corresponds to a distillation column in the actual flow, there are two branches-right branch and left branch for the node. The right branch corresponds to the top stream of the column that includes the components lighter than the light key component while the left branch corresponds to the bottom stream that includes the components heavier than the heavy key component. If a splitter node is selected as the root node for a GP code, two branches (left branch and right branch) are also generated which correspond to the same compositions and physical properties but different flow rates of the two outlet streams for this splitter, respectively. Once the root node for GP tree is confirmed, no matter if it is a column node or a splitter node, the connection relation for both the left branch and right branch with the following nodes should be expressed. It is specified in the algorithm that the next subnode can be randomly selected from three types of nodes, including column node, splitter node, and blender node. Then, every subnode connects to the next subnode produced randomly, that results in each branch of any node extending downward continuously. The connection proceeds alternately until all branches reach the GP code tree’s maximum deepness which has been defined in this paper. Then all branches of the GP code are randomly connected

to the different terminal product blenders to produce the specified product streams. Now, the growth of the GP tree stops; that means a whole GP tree-like code has been achieved (Figure 4). During the GP-code forming process, if any branch of any subnode is confirmed randomly to be connected to a blender node, a pair of property parameters of this blender node can be used to indicate the position of this blender node, that is, this blender locates in which layer and which column among the blender grid framework. It means that any node of any subtree in GP code has the possibility of connecting to any blender node in the blender grid structure. In other words, any blender has the possibility to mix multiple streams from different subtrees. Then, this blender node will be as a subroot node, and a corresponding subtree downward grows randomly based on it. The coding method can be used to express the bypass operation successfully. When a splitter node is confirmed as the root node of a GP tree-like code randomly and a terminal products blender of the GP code is just selected by one of the two branches of this splitter node directly, it means that a part of the original feed can be sent to products blender directly without being separated by a distillation column. This connecting way can be used to express the bypass operation for original feed stream. At the same time, for any subnode locating in anywhere of the whole GP tree-like code, when one branch of it is connected to the terminal products blender directly, it also means that a part of the inlet stream for this node can be sent to the terminal products blender directly without being separated by a distillation column, and the bypassing operation for this middle stream can be realized. 3.3. GP Evolutionary Operations Based on Chemical Engineering Domain Knowledge. Similar to GA, GP also needs the operations of reproduction, crossover, and mutation to achieve evolutionary optimization process. 3.3.1. Reproduction Operation. A reproduction operation is used to regenerate the optimal members from the population of one generation to the next generation and eliminate the bad members. The operation step for athletics selection method11 is followed: first, a small population including some individuals is selected randomly from the total population and the fitness value of every individual is calculated. Then, the individual that has the optimal fitness value is reproduced. The above operation proceeds repeatedly and some excellent individuals are produced by reproduction operation. Because this method does not need to evaluate the fitness of every member in one generation, which speeds up the evolutionary velocity of the population obviously, it is adopted to execute the reproduction operation in this paper. 3.3.2. Crossover Operation. Crossover operation is used to exchange the part genes of two members of GP code to produce two new members, and it is divided into two classes in this work, that is, the crossover of node and the crossover of node property. When the crossover of node proceeds, according to the special feature of multicomponent products nonsharp distillation sequence, it is specified in this work that two column nodes can exchange with each other or two splitter nodes can exchange with each other. Also a column node and a splitter node can exchange with each other too, but a column node or a splitter node cannot be crossed to a blender node and two blender nodes cannot be crossed to each other also. If two allowed nodes will be crossed, in order to guarantee the validity of this operation, it is specified that the two nodes must have the same feed composition. The judgment method is to compare whether the

Ind. Eng. Chem. Res., Vol. 47, No. 22, 2008 8819 -1

Table 2. Flow Rates for Example 1 (kmol · h ) A propane

B isobutane

C n-butane

D isopentane

E n-pentane

quantity

45.36 20 25.36

136.08 100 36.08

226.8 100 126.8

181.44 150 31.44

317.52 100 217.52

907.2 470 437.2

feed product 1 product 2

Table 3. The Results of the Optimal Configuration for Example 1a tower number

D (m)

Nt

Ft

P (atm)

R

Qc (106 kJ/h)

Qr (106 kJ/h)

1 2 3 4 5

0.74 1.26 0.47 0.19 0.09

58 75 23 22 22

24 37 13 9 9

9.6 3.8 4.9 15.1 15.1

8.1 5.8 1.7 3.3 3.3

3.46 7.51 1.16 0.25 0.053

4.03 8.13 1.26 0.27 0.057

a Annual operation cost ) $339,889; annual equipment cost ) $87,686; annual total cost ) $427,576.

Figure 6. (a) GP code for Example 1 (numbers are streamflow rates); (b) the optimal separation configuration for example 1.

feed composition information contained in the two nodes is the same or not. When the crossover of the node proceeds, the contents of the crossover operation include not only the node

itself but also all properties of the node. At the same time, both the node itself and the two branches of this node partake in the crossover operation. Here, every blender node cannot partake in the crossover operation in order to protect the blender grid frame. Figure 5 gives an example that expresses the process of crossover operation. There are two nodes that are circled with a broken line in code 1 and 2. If they have the same feed composition information, they can cross the node itself and the branches under each other and two new GP codes are produced. When the crossover operation aims at node property, it is defined that only corresponding properties of the two homogeneous nodes are crossed while the node itself remains changeless. Similarly, every blender node cannot partake in the crossover of node property in order to protect the blender grid frame. To ensure the randomization of this algorithm, stochastic selection is adopted in this paper to determine the types of the crossover operation (i.e., the crossover of node or the crossover of node property). 3.3.3. Mutation Operation. According to the characteristic of a nonsharp distillation process, it is specified in this work that the column node itself and splitter node itself can participate in the mutation operation and they can mutate to each other too. When the mutation operation is for the column node or splitter node itself, it is limited in that the mutation operation cannot change the feed information for the node but can mutate for the information of stream cut. The mutation operation method for column node or splitter node is that the left and right subnodes of the selected node are deleted first. Then, the information of stream cut is randomly changed again in the range of the feed information included in this node. Lastly, the new left and right branches of this node are generated according to the new information of stream cut that can ensure the mutated code is valid. Here, the new information of stream cut for the column node is to assign the light and heavy key components randomly again and for the splitter node the new information of stream cut is to generate the distribution coefficient randomly again. When the column node and the splitter node mutate to each other, the feed information for every node must be kept fixed first, and the new left and right branch will be produced according to the new stream cut information (or distribution coefficient) of this new node. In the same way, the types of mutation operation (i.e., the mutation operation for homogeneous node or the mutation operation for different type node) is randomly specified here in order to ensure the diversity of the GP code. In summary, the GP tree-like code can be used to represent every individual of the population directly, and the evolution operations can proceed in the code forthright, not like the GA which uses the binary character to represent separation flow. So, the improved GP-based algorithm can be used to describe various netlike structures of nonsharp separation process directly and may discover more unimaginable flow structures. 4. The Cost Model According to the definition for GP encoding method and evolutionary operations from above, the algorithm can generate

8820 Ind. Eng. Chem. Res., Vol. 47, No. 22, 2008 Table 4. Flow Rate for Example 2 (kmol · h-1) A propane

B iso-butane

C n-butane

D iso-pentane

E n-pentane

F n-hexane

G n-heptane

quantity

15 7 3 5

20 4 8 8

25 10 7 8

35 15 10 10

30 12 9.4 8.6

45 20 15 10

30 15 10 5

200 83 62.4 54.6

feed product 1 product 2 product 3

Table 5. The Results of the Optimal Configuration for Example 2a tower number D (m) Nt 1 2 3 4 5 6

0.12 0.12 0.14 0.04 0.12 0.16

15 68 76 15 61 89

Ft P (atm) 7 28 46 8 31 43

12.52 9.40 10.83 13.56 14.44 17.32

R 0.98 10.99 3.54 0.91 3.09 13.96

Qc (106 kJ/h) Qr (106 kJ/h) 0.10 0.10 0.12 0.01 0.09 0.18

0.14 0.12 0.13 0.01 0.11 0.19

a Annual operation cost ) $17,143; annual equipment cost ) $16,157; annual total cost ) $33,300.

0.1. To improve the accuracy of the computation, the maximum and minimum values of column operation pressure, the top and bottom temperatures of column and the enthalpies of the feed, and top and bottom stream of the column are all calculated with the rigorous PR equation of states. When the area of heat exchanger is calculated, the heat transfer coefficients are set as 780 and 950 w/(m2 · k), which partly correspond to condenser and reboiler. The condensers and the reboilers require utilities to perform heat exchange. The low grade heat vapor is chosen first in the precondition of satisfying minimal heat transfer difference when the exterior heat is selected as heat source. The computation is performed using the grade and price of utilities as illustrated in Table 1. 4.2. Calculation of Objective Function. Fitness is the principal criterion to judge whether the individual is good or not, and the annual total cost (including equipment cost and operation cost) is adopted as the fitness in this paper. It is easy to conclude that the smaller the fitness is, the better the code is. Calculation of the fitness includes the following steps: (1) convert every GP code generated as above-mentioned to the corresponding nonsharp separation flow; (2) design every distillation column of the flow to obtain equipment parameters and operation parameters; (3) the equipment costs and the operation costs are calculated to obtain the fitness for every GP code. 5. Calculation Steps for the GP-Based Synthesis Algorithm

Figure 7. The optimal separation configuration for example 2.

numerous tree codes corresponding to various actual separation structures. The next step is to optimize various parameters of the separation system such as operation pressure, etc., then the column diameters and column sizing of all columns for every flow are confirmed, which can be used to calculate the total cost of every nonsharp separation flow, with an aim to find the optimal flow with the minimum cost. 4.1. Column and Heat Exchanger Design. We adopt the traditional Fenske-Underwood-Gilliland abridged method for the design calculations of column equipment, and the Guthrie relationship22 is introduced for the computation of equipment cost with the equipment annual depreciation coefficient being

The GP-based synthesis algorithm includes the following steps: (1) Generate initial population randomly as described in 3.2. Because too many nodes and blender layers can lead to an excessively complex GP code, even to the point of making the calculation of this algorithm fail, the sum of total nodes is limited to ensure the successful execution of this algorithm. In this paper, the maximal depth of the GP algorithm tree and the maximal blender layer number that includes the last layer of the terminal product blender will be restricted, while the size of population is set to V and the maximal evolutionary generation is set to W, where V and W will be assigned according to every actual problem. (2) The complex algorithm is used to optimize the continuous variablesthe distribution coefficient of every splitter and the operation pressure of every distillation column for every GP code, while the GP algorithm is used to optimize discrete variablessnonsharp separation structure. That is to say, the GPbased strategy improved in this paper can deal with both discrete and continuous variables simultaneously. (3) Translate every GP code into its corresponding nonsharp separation flow and get the annual total cost for it. (4) The operations of reproduction, crossover, and mutation are performed as described in 3.3, where the reproduction rate, crossover rate, and mutation rate will be set according to the specific example.

Ind. Eng. Chem. Res., Vol. 47, No. 22, 2008 8821 -1

Table 6. Flow Rate for Example 3 (kmol · h )

feed product product product product product

1 2 3 4 5

A methanol

B ethanol

C ipropanol

D npropanol

E ibutanol

F nbutanol

G npentanol

H 1-hexanol

I C-hexanol

J 1-heptanol

K 1-octanol

quantity

10 0.05 0 9.95 0 0

20 0.1 0 14.4 0 5.5

5 0 0 3.6 0 1.4

10 0 4.5 0.3 0 5.2

5 0 1 0 1.1 2.9

15 0 3 0 3.2 8.8

5 0 1 0 1.1 2.9

10 0 2 0 8 0

5 0.33 1 0 3.67 0

5 0.4 4.6 0 0 0

10 0.75 9.25 0 0 0

100 1.63 26.35 28.25 17.07 26.7

Table 7. the Results of the Optimal Configuration for Example 3a tower number D (m) 1 2 3 4 5 6 7 8 9 10

0.25 0.04 0.19 0.14 0.10 0.06 0.02 0.18 0.07 0.06

Nt

Ft

P (atm)

R

24 20 29 20 89 20 35 25 22 58

15 10 17 11 52 14 15 14 10 38

1.23 2.21 1.45 2.32 1.69 1.98 2.45 1.60 1.75 2.15

0.07 2.02 0.47 0.29 9.10 0.72 7.27 0.69 3.06 4.41

Qc (106 kJ/h) Qr (106 kJ/h) 0.71 0.02 0.46 0.19 0.10 0.04 0.004 0.40 0.08 0.05

1.00 0.02 0.52 0.24 0.10 0.04 0.004 0.42 0.08 0.05

a Annual operation cost ) $60,678; annual equipment cost ) $24,396; annual total cost ) $85,074.

(5) Steps 2-4 are repeated until the fitness does not change in the last two generations or the number of the maximal generation is reached, which means the algorithm should be terminated. 6. Examples The calculation software for the GP-based synthesis algorithm is programmed using C++ language and the effectiveness of the algorithm is validated by the following three typical nonsharp distillation problems on a Pentium 4/3.0 G computer. The optimal nonsharp separation flow will be searched, which can minimize the annual total cost.

Figure 8. The optimal separation configuration for example 3.

Example 1 is the second hydrocarbon mixture example of Liu and Xu14 and the data are given in Table 2. Through a lot of tests, the size of population for GP is set to 60 and the reproduction ratio, crossover ratio, and mutation ratio is set to 10, 50, and 5%. Because Liu and Xu14 did not study the nonsharp separation problem based on the annual total cost (including equipment cost and operation cost), they did not give the optimal structural decision flow and did not give the related parameters D, Nt, Ft, etc. In this paper, annually total cost, for the first time, is used as an optimization objective for the nonsharp separation system, and the equipment design and cost model is proposed. The optimal GP code and the corresponding optimal separation configuration obtained by this work is expressed in Figure 6a,b, which is obtained at the evolutional generation of 12, and the time for CPU use is about 300 s. The results for the optimal configuration are listed in Table 3 with the related parameters, D (distillation tower diameter), Nt (number of plate), Ft (position of feed), P (pressure), R (reflux ratio), Qc (thermal load of condenser), and Qr (thermal load of reboiler). Example 2 is a seven-component three-product hydrocarbon mixture nonsharp separation problem proposed by this paper and the data are given in Table 4. Because no seven-component nonsharp separation problem has been reported, no valuable data can be used to compare with this work. The size of the population for GP is set to 80, and the reproduction ratio, crossover ratio, and mutation ratio is set to 15, 60, and 8%,

8822 Ind. Eng. Chem. Res., Vol. 47, No. 22, 2008

respectively. The optimal separation configuration obtained by this work is expressed in Figure 7, which was obtained at the evolutional generation of 20 and a CPU time of ca. 600 s. The results for the optimal configuration are listed in Table 5. Example 3 is an 11-component 5-product alcohol mixture nonsharp separation problem proposed by this paper, and the data are given in Table 6. The size of population for GP is set to 100 and the reproduction ratio, crossover ratio, and mutation ratio is set to 20, 60, and 10%, respectively. The optimal separation configuration found by this work is expressed in Figure 8. It is difficult to find such a complex flow with the algorithm of experimental rules, and it cannot be included in the superstructure of the algorithm of mathematical programming either. However, the GP-based synthesis algorithm in this work can get the optimal solution at the evolutional generation of 25 within about 800 s without being given any superstructure and clue information. The results for the optimal configuration are listed in Table 7. 7. Conclusion A new stochastic optimization algorithmsgenetic programming (GP) is proposed for the synthesis of multicomponent products nonsharp distillation sequences. Based on the special blender grid framework, the complex nonsharp distillation netlike structure is devised and can be expressed directly using GP’s special hierarchical configuration. In this paper, the annual total cost, for the first time, is used as an optimization objective, and the equipment design and cost model of nonsharp separation system is given. To ensure the proceeding validity of this algorithm, several evolutionary factors are improved according to the domain knowledge of chemical engineering. While GP is used to optimize the structural variable, the complex algorithm is used to optimize the continuous variable simultaneously. Typical nonsharp separation problems that include hydrocarbon and alcohol mixtures, for the first time, are solved using the above strategy. A bigger size of example, as much as 11 components, is solved and compared with the reported 5 components example. It is shown by the computation results that the proposed GP-based synthesis method can automatically solve the optimization synthesis problem of the nonsharp distillation sequence including distillation, splitting, blending, and bypassing operations quickly and effectively. Nomenclature D ) distillation tower diameter, m Ei ) the ith component Ft ) position of feed N ) component number for mixture Nt ) number of plate P ) pressure, atm Qc ) thermal load of condenser, kJ · h-1 QR ) thermal load of reboiler, kJ · h-1 R ) reflux ratio V ) size of the population for GP W ) maximal number of evolutionary generation for GP

Literature Cited (1) Agrawal, R. More operable fully thermally coupled distillation column configurations for multicomponent distillation. Trans. Inst. Chem. Eng. 1999, 77, 543–553. (2) Agrawal, R. Thermally coupled distillation with reduced number of intercolumn vapor transfers. AIChE J. 2000, 46, 2198–2210. (3) Agrawal, R. A method to draw fully thermally coupled distillation column configurations for multicomponent distillation. Trans. Inst. Chem. Eng. 2000, 78, 454–461. (4) Caballero, J. A.; Grossmann, I. E. Generalized disjunctive programming model for the optimal synthesis of thermally linked distillation columns. Ind. Eng. Chem. Res. 2001, 40, 2260–2274. (5) Grossmann, I. E. Review of nonlinear mixed-integer and disjunctive programming techniques. Optimiz. Eng. 2002, 3, 227–252. (6) Ismail, S. R.; Proios, P.; Pistikopoulos, E. N. Modular synthesis framework for combined separation/reactive systems. AIChE J. 2001, 47, 629–649. (7) Grossmann, I. E.; Aguirre, Pio, A.; Bartteld, M. Optimal synthesis of complex distillation columns using rigorous models. Comput. Chem. Eng. 2005, 29, 1203–1215. (8) Caballero, J. A.; Grossmann, I. E. Structural considerations and modeling in the synthesis of heat-integrated-thermally coupled distillation sequences. Ind. Eng. Chem. Res. 2006, 45, 8454–8474. (9) Nath, R. Studies in the synthesis of separation process. Ph.D. Dissertation. University of Houston, TX, 1977. (10) Muraki, M.; Hayakawa, T. Synthesis of multicomponent multiproduct separation process with nonsharp separators. Chem. Eng. Sci. 1988, 43, 259–268. (11) Liu, Z.; Guo, Z.; Chen, Z. Synthesis of distillation sequences with multi-feed by using an individual material allocation diagram. J. Chem. Ind. Eng. (China) 1994, 45, 321–326. (12) Bamopoulos, G.; Nath, R.; Motard, R. L. Heuristic synthesis of nonsharp separation sequences. A.I.Ch.E.J. 1988, 34, 763–780. (13) Cheng, S. H.; Liu, Y. A. Studies in chemical process design and synthesis 8. A simple heuristic method for systematic synthesis of initial sequences for sloppy multicomponent separations. Ind. Eng. Chem. Res. 1988, 27, 2304–2322. (14) Liu, Z.; Xu, X. Heuristic procedure for the synthesis of distillation sequences with multicomponent products. Chem. Eng. Sci. 1995, 50, 1997– 2004. (15) Liu, Z.; Hu, L.; Jin, Z.; Xien, X. Two-stage procedure for the synthesis of nonsharp distillation flowsheets. J. Chem. Eng. Chin. UniV. 1999, 13, 66–72. (16) Hu, Z.; Chen, B.; He, X. Heuristic synthesis of multicomponent products with simple and sharp separators. Comput. Chem. Eng. 1993, 17, 379–397. (17) Wehe, R. R.; Westerberg, A. W. An algorithmic procedure for the synthesis of distillation sequences with bypass. Comput. Chem. Eng. 1987, 11, 619–627. (18) Aggarwal, A.; Floudas, C. A. Synthesis of heat integrated nonsharp distillation sequences. Comput. Chem. Eng. 1992, 16, 89–108. (19) Zhang, L.; Linninger, A. A. Towards Computer-aided Separation Synthesis. AIChE J. 2006, 52, 1392–1409. (20) Leboreiro, J.; Acevedo, J. Processes synthesis and design of distillation sequences using modular simulators: a genetic algorithm framework. Comput. Chem. Eng. 2004, 28, 1223–1236. (21) Koza, J. R. Genetic Programming: on the Programming of Computer by Means of Natural Selection; The MIT Press: Cambridge, MA, 1992. (22) Douglas, J. M. Conceptual Design of Chemical Processes; Chemical Engineering Series; McGraw-Hill: New York, 1988.

ReceiVed for reView April 15, 2008 ReVised manuscript receiVed July 9, 2008 Accepted September 5, 2008 IE800610S