Science

Computers and Chemical Engineering 27 (2003) 647 /668 www.elsevier.
com/locate/compchemeng
New algorithms for mixed-integer dynamic optimization

Vikrant Bansal 1, Vassilis Sakizlis, Roderick Ross 2, John D. Perkins, Efstratios N. Pistikopoulos *
Centre for Process Systems Engineering, Department of Chemical Engineering, Imperial College, Prince Consort Road, London SW7 2BY, UK Received 19 February 2001; received in revised form 21 October 2002; accepted 11 November 2002
Abstract Mixed-integer dynamic optimization (MIDO) problems arise in chemical engineering whenever discrete and continuous decisions are to be made for a system described by a transient model. Areas of application include integrated design and control, synthesis of reactor networks, reduction of kinetic mechanisms and optimization of hybrid systems. This article presents new formulations and algorithms for solving MIDO problems. The algorithms are based on decomposition into primal, dynamic optimization and master, mixed-integer linear programming sub-problems. They do not depend on the use of a particular primal dynamic optimization method and they do not require the solution of an intermediate adjoint problem for constructing the master problem, even when the integer variables appear explicitly in the differential /algebraic equation system. The practical potential of the algorithms is demonstrated with two distillation design and control optimization examples. # 2002 Elsevier Science Ltd. All rights reserved.
Keywords: Mixed-integer optimization; Dynamic modelling; Process design; Process control; Distillation; gPROMS
1. Introduction The last decade has seen a rapid increase in the use of dynamic simulation as a tool for studying the transient behaviour of process systems. In fact, common practice is to use dynamic simulation as a means towards optimizing process design and/or operation. This is achieved by an engineer performing multiple, trial-anderror simulations in order to choose a set of values for the process parameters that leads to good performance (e.g. low cost or fast transition between steady-states) while satisfying the constraints under which the process operates. Although this approach is simple and intuitive, it has the major drawback that the fraction of the parameter space that can be searched is severely limited, especially as the number of parameters increases. This not only means that it is impossible to guarantee
* Corresponding author. Tel.: '/44-20-7594-6620; fax: '/44-207594-6606. E-mail address: e.pistikopoulos@ic.ac.uk (E.N. Pistikopoulos). 1 Present address: Orbis Investment Advisory Limited, Orbis House, 5 Mansfield Street, London W1G 9NG, UK. 2 Present address: United Technologies Research Center, 411 Silver Lane, East Hartford, CT 06108, USA.
optimality, but in many cases it can even be difficult to find a set of parameter values that leads to a feasible solution where all the constraints are satisfied (Pantelides, 1996). The ad hoc approach described above can be avoided by posing formal, dynamic optimization problems, where the objective is to select the set of time-dependent and/or time-invariant parameters that optimize some aspect of the transient behaviour of the process while satisfying a given set of constraints. A number of algorithms have been developed in the literature for solving such problems and their reliability has evolved to the extent that realistic engineering problems involving thousands of variables can now be readily solved with commercial codes such as gPROMS/gOPT (Process Systems Enterprise Ltd., 2000). Reported applications using gPROMS include optimization of operating policies in multi-component batch distillation and other processes (Charalambides, Shah, & Pantelides, 1995; Furlonge, Pantelides, & Sorensen, 1999; Ishikawa, Natori, Liberis, & Pantelides, 1997); operation of catalytic reactors (Lovik, Hillestad & Hertzberg, 1998); design and operation of heat exchangers under fouling (Georgiadis, Rotstein, & Macchietto, 1998); analysis of hazards in
0098-1354/02/$ - see front matter # 2002 Elsevier Science Ltd. All rights reserved. doi:10.1016/S0098-1354(02)00261-2
648
V. Bansal et al. / Computers and Chemical Engineering 27 (2003) 647 /668
batch processes (Viswanathan, Shah, & Venkatasubramanian, 2000); and the simultaneous optimization of the process and control system design in complex distillation systems (Bansal, Ross, Perkins, & Pistikopoulos, 2000; Ross, Bansal, Perkins, Pistikopoulos, Koot, & van Schijndel, 1999). Other notable applications of dynamic optimization to chemical engineering processes include on-line model predictive control (Biegler & Sentoni, 2000; Nath & Alzein, 2000); fault diagnosis (Huang, Reklaitis, & Venkatasubramanian, 2000); determination of policies for switching between steady-states (Cervantes, Tonelli, Brandolin, Bandoni & Biegler, 2000; Mcauley & Macgregor, 1992; White, Perkins, & Espie, 1996); and optimization of systems subject to exceptional events such as failures and sudden changes in product demand (Abel & Marquardt, 2000). Practical engineering problems generally involve discrete as well as continuous decisions. For example, in the optimal design of a distillation column, the number of trays and feed location take discrete values, while the values of the column diameter and surface areas of the reboiler and condenser are continuous. This gives rise to mixed-integer dynamic optimization (MIDO) problems, the solution of which is a formidable task. The development of algorithms for solving such problems is still in its infancy and thus few large-scale applications have been reported. The objective of this article is to propose new algorithms for solving MIDO problems that do not depend on particular integration and dynamic optimization solution methods, require significantly less intermediate information than current algorithms, and as such, are better suited to the solution of larger, more realistic problems. The remainder of the article is structured as follows. After some mathematical preliminaries, the existing solution approaches for MIDO are thoroughly and critically reviewed. An alternative formulation and algorithm based on generalised Benders decomposition principles is then presented and its advantages are discussed. An analogous algorithm based on outerapproximation (OA) is also outlined and, finally, the application of the two algorithms is illustrated with two distillation design and control examples.
hp (xd (ti ); xd (ti ); xa (ti ); u(ti ); d; y; ti )00; i 0 1; . . . ; N; gp (xd (ti ); xd (ti ); xa (ti ); u(ti ); d; y; ti )00; i 0 1; . . . ; N; hq (d; y)00; gq (d; y)00;
ti [t0 ; tf ];
ti [t0 ; tf ];
(1)
where hd 0/0 and ha 0/0 represent the system of differential /algebraic equations (DAEs) with initial conditions h0 0/0; hp 0/0 and gp 0/0 represent the set of point equalities and inequalities, respectively, that must be satisfied at specific time instances (including the end-point if required); hq 0/0 and gp 0/0 are the timeinvariant equality and inequality constraints; xd(t) and xa(t) are the vectors of differential state and algebraic variables, respectively; u(t) is the vector of time-varying control variables; d is the vector of time-invariant continuous search variables (that may include the final time, tf); and y is the vector of time-invariant integer variables. In the sequel, we will consider y to be a vector of binary (0 /1) variables since most process engineering applications are modelled in this way and any integer variable can be expressed as a sum of binary variables (Floudas, 1995). The case of time-varying binary variables will be considered later in the article. Note that the formulation in Eq. (1) is quite general. For example, integral terms in the objective function can be eliminated by the addition of extra differential equations. If there are path inequality constraints that must be satisfied at all times during the time horizon, these can always be converted into end-point inequality constraints (e.g. Bansal, 2000; Sullivan & Sargent, 1979), which are a sub-set of gp 0/0.
3. Review of existing solution approaches for MIDO By definition, it is the presence of the binary variables y that complicate the solution of a MIDO problem compared to a continuous dynamic optimization problem (for which solution techniques are well-established). In the context of optimally designing dynamic systems for robustness, Samsatli, Papageorgiou and Shah (1998) proposed the use of smooth approximations for the binary variables in order to convert the discrete / continuous problem Eq. (1) into a purely continuous one. For example, if a binary variable y is defined such that it takes a value of 1 if another (continuous) variable x /0 and is zero otherwise, then the following smooth approximation would be used: 1 y 0 [tanh(bx)'1]; 2 (2)
2. Preliminaries A MIDO problem can be formulated as:

u(t);d;y
min J(xd (tf ); xd (tf ); xa (tf ); u(tf ); d; y; tf );
s:t: hd (xd (t); xd (t); xa (t); u(t); d; y; t)00; t [t0 ; tf ]; ha (xd (t); xa (t); u(t); d; y; t)00; t [t0 ; tf ]; h0 (xd (t0 ); xd (t0 ); xa (t0 ); u(t0 ); d; y; t0 )00;
649
where b is a large, positive number. The major drawback of trying to solve general MIDO problems using such an approach is that the smooth approximation does not always lead to integral values of y (e.g., in Eq. (2), when x 0/0, y0/0.5). Non-integral values of y will often render the engineering problem physically meaningless and lead to severe numerical difficulties. Androulakis (2000) proposed a branch-and-bound algorithm for solving MIDO problems that arise when attempting to systematically reduce the complexity of kinetic mechanisms. This also requires the solution of dynamic optimization problems with relaxed binaries y. Furthermore, as acknowledged by the author, branchand-bound techniques are unsuitable for all but very small MIDO problems since they will typically involve the solution of a much larger number of dynamic optimization problems (the computational rate-determining step) than the decomposition methods described below. The other MIDO solution approaches that have appeared in the literature are all based on decomposition principles, but differ in their treatment of the DAE system. One approach is to completely discretise the system using a technique such as orthogonal collocation on finite elements. This converts the MIDO problem Eq. (1) into a finite-dimensional, mixed-integer non-linear program (MINLP), where the continuous search variables for the optimization are the full set of discretisation parameters for the differential, algebraic and control variables, and the time-invariant variables d. The MINLP can be solved using any conventional method. Balakrishna and Biegler (1993) and Avraam, Shah and Pantelides (1998, 1999) used the outerapproximation/equality-relaxation/augmented-penalty (OA/ER/AP) algorithm of Viswanathan and Grossmann (1990), while Dimitriadis and Pistikopoulos (1995) and Mohideen, Perkins and Pistikopoulos (1996) used generalised Benders decomposition, or GBD (Geoffrion, 1972), in the contexts of synthesising reaction/separation systems, optimizing hybrid processes, analysing dynamic systems under uncertainty, and simultaneous design and control, respectively. The advantages of this approach are that the dynamic model and optimizer constraints are simultaneously converged and that general types of inequality constraints are easily handled. There are two major drawbacks, however. Firstly, the MINLP resulting from the discretisation is very large even for relatively small problems. In the context of solving realistic engineering problems the size of the MINLP will often make the problem intractable. Secondly, intermediate solutions during the course of the optimization generally do not satisfy the original differential /algebraic system equations (hence the common use of the term infeasible path methods). This effectively renders the approach an allor-nothing strategy since if the optimizer fails to find an
optimum, the intermediate results are usually meaningless and cannot be used, e.g., as new initial starting points. For continuous dynamic optimization problems, an alternative approach to complete discretisation is to only parameterise the control variables, u(t ), in terms of time-invariant parameters (reduced space discretisation or control vector parameterisation). For given u(t) and values of the other search variables, d, the DAE system is integrated using standard algorithms. Gradient information with respect to the objective function and constraints is obtained through finite difference perturbations, or more accurately through integration of the sensitivity equations (e.g. Vassiliadis, Sargent & Pantelides, 1994a,b), or solution of the adjoint equations (e.g. Sullivan & Sargent, 1979; Pytlak, 1998; Bloss, Biegler, & Schiesser, 1999). An NLP solver then adjusts the values of the controls and other search variables on the basis of these gradients. Although this method has the disadvantage that the solution of the DAEs and other equations can be computationally expensive, it gives useful intermediate (sub-optimal) solutions that are feasible with respect to the DAE system; allows more accurate control of the error associated with the solution of the DAE system; and is better conditioned when handling stiff models. As such, it is much more suited for large dynamic systems than the complete discretisation approach, although work is ongoing aimed at redressing the balance (e.g. Cervantes & Biegler, 1998). The MIDO algorithms in the literature that utilise reduced space methods (Allgor & Barton, 1999; Mohideen, Pistikopoulos & Perkins, 1997; Narraway, 1992; Schweiger & Floudas, 1997; Sharif, Shah & Pantelides, 1998) all decompose the problem into a series of primal problems where the binary variables are fixed, and master problems which determine a new binary configuration for the next primal problem. In all cases the primal problem corresponds to a dynamic optimization problem that is solved using a reduced space approach, and which gives an upper bound on the final solution. The methods mainly differ in how they construct the master problem, which gives a lower bound on the solution. In the case of OA/ER/AP, the MILP master problem is constructed by linearising the objective function and constraints. This contrasts with GBD where the master problem is constructed using the solution of the primal problem and corresponding dual information. The advantage of the former approach is that it requires the solution of fewer primal problems since its (more constrained) master problem always gives at least as tight a lower bound as the GBD master problem (Duran & Grossmann, 1986). This could be important for MIDO problems since the primal problem (dynamic optimization) will tend to be much more computationally expensive relative to the master problem than for MINLPs. However, the OA methods
650
are restricted to problems where the binary variables appear linearly and separably in the objective function and constraints, and are not designed to handle binary variables that appear explicitly in the DAE system (pointed out by Schweiger & Floudas, 1997), which would be the case, e.g., for the simultaneous design and control optimization of a distillation column. Note that the convexity conditions required for OA/ER/AP and GBD to give the global optimum are rarely satisfied in dynamic process engineering problems. Allgor and Barton (1999) proposed a different master problem approach where an MILP is solved based on a so-called screening model. This approach may have the potential for providing rigorous bounds on the global optimum; its major drawback, however, is that the screening model is entirely case study-specific since it is derived from domain specific knowledge gathered from physical laws and engineering insight . Mohideen et al., 1997 and Schweiger and Floudas (1997) have both developed MIDO algorithms based on variant-2 of generalised Benders decomposition, v2GBD (Floudas, 1995), which are now explored in more detail. In both cases, the k th primal problem takes the form: min J(xd (tf ); xd (tf ); xa (tf ); u(tf ); d; y ; tf );
u(t);d k
The master problem is formulated using dual information and the solution of the primal problem in order k to construct supporting functions about uk (t) and d : If the primal problem is solved using adjoint-based arguments, as in Mohideen et al. (1997), then the dual variables are available directly from the gradient evaluation method used for the primal problem. If a sensitivity-based approach is used for the primal problem then the dual variables are obtained through the solution of the final-value problem (Bryson & Ho, 1975; Vassiliadis, 1993): d @hd T k @hd T k @ha T k nd (t) ( nd (t)( na (t)00; dt @ xd @xd @xd @hd T k @ha T k ( nd (t)( na (t)00; @xa @xa with boundary conditions: @J T @hd T k @hd T k ' nd (tf )' jd;f @xd @ xd @xd tf tf @ha T k ' ja;f @xd tf (4)
s:t: hd (xd (t); xd (t); xa (t); u(t); d; yk ; t)00; t [t0 ; tf ]; ha (xd (t); xa (t); u(t); d; yk ; t)00; t [t0 ; tf ]; h0 (xd (t0 ); xd (t0 ); xa (t0 ); u(t0 ); d; yk ; t0 )00; hp (xd (ti ); xd (ti ); xa (ti ); u(ti ); d; yk ; ti )0 0; i 01; . . . ; N; gp (xd (ti ); xd (ti ); xa (ti ); u(ti ); d; yk ; ti )0 0; i 01; . . . ; N; hq (d; yk )0 0; gq (d; yk )0 0; (3) ti [t0 ; tf ]; ti [t0 ; tf ];
@hp T k @gp T k ' mp ' lp 00; @xd @xd
(5)
where nd, na, m and l are the dual multipliers associated with the differential equations, algebraic equations, equality point constraints and inequality point constraints, respectively. jd, f and ja, f are multipliers that are associated with the differential and algebraic equations at the final time; these are evaluated from the first order optimality conditions of the primal problem (Vassiliadis, 1993). The dual multipliers mk and lk are obtained from the solution of the primal problem, regardless of the method used to solve it. The Lagrange function, Lk, for the case of a feasible primal problem Eq. (3), is given by: Lk (uk ; d ; y)0J(xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; y; tf ) d a '
k k k
where Eq. (3) corresponds to Eq. (1) with the binary variables fixed, y0/yk . In the approach of Mohideen et al. (1997), the primal problem is solved using an implicit Runge /Kutta method to integrate the DAE system and the NLP gradient information is obtained using adjointbased arguments. In the approach of Schweiger and Floudas (1997), a backward difference formula algorithm is used for the DAEs and the sensitivity equations are integrated to obtain the required gradients. The solution of the k th primal problem is denoted as uk (t) k and d ; with resulting differential variables xk (t) and d algebraic variables xk (t):/ a
g g
tf t0 tf t0
T
T k k nk (t)hd (xd (t); xk (t); xk (t); uk (t); d ; y; t)dt d a d
'
nk (t)ha (xk (t); xk (t); uk (t); d ; y; t)dt d a a

k k
'jk hd (xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; y; tf ) d a d;f 'jk ha (xk (tf ); xk (tf ); uk (tf ); d ; y; tf ) d a a;f d a 'jk hd (xd (t0 ); xk (t0 ); xk (t0 ); uk (t0 ); d ; y; t0 ) d;0
T T

T k 'jk ha (xk (t0 ); xk (t0 ); uk (t0 ); d ; y; t0 ) d a a;0
651
'kk h0 (xd (t0 ); xk (t0 ); xk (t0 ); uk (t0 ); d ; y; t0 ) d a

T k k 'mk hp (xd (ti ); xk (ti ); xk (ti ); uk (ti ); d ; y; ti ) d a p
(7). In order to achieve this, a modification of problem Eq. (1) is considered where the original binary variables, y, are relaxed, but are still forced to take integral values by the addition of new, binary variables, y; and constraints hy 0 y( y00:/ 4.1. Primal problem (6) At the k th iteration of the GBD algorithm, the following dynamic optimization problem is solved:
u(t);d;y
'lk gp (xd (ti ); xk (ti ); xk (ti ); uk (ti ); d ; y; ti ) d a p 'mk hq (d ; y)'lk gq (d ; y); q q

T
where k is the set of multipliers associated with the initial conditions, and jd, 0 and ja, 0 are the sets of multipliers associated with the initial values of the DAEs. These are all evaluated in a similar manner to jd, f and ja, f. The relaxed master problem is then: min h;
y;h k s:t: Lk (uk ; d ; y)0 h;
min J(xd (tf ); xd (tf ); xa (tf ); u(tf ); d; y; tf );
hd (xd (t); xd (t); xa (t); u(t); d; y; t)0 0; t [t0 ; tf ]; ha (xd (t); xa (t); u(t); d; y; t)00; t [t0 ; tf ]; h0 (xd (t0 ); xd (t0 ); xa (t0 ); u(t0 ); d; y; t0 )0 0;
k Kfeas ;
(7)
hp (xd (ti ); xd (ti ); xa (ti ); u(ti ); d; y; ti )00; i 0 1; . . . ; N; gp (xd (ti ); xd (ti ); xa (ti ); u(ti ); d; y; ti )00; i 0 1; . . . ; N; hq (d; y)00; gq (d; y)00; hy 0y( yk 00;
ti [t0 ; tf ];
where Kfeas is the set of all feasible primal problems solved up to the iteration under consideration. If the variables y participate linearly in Eq. (1), then Eq. (7) is an MILP that can be solved via conventional branch and bound methods. Note that integer cuts of the form (Floudas, 1995): X X yi ( yi 0jBk j(1;
i Bk i NBk
ti [t0 ; tf ];
(9)
0 fi:yk 01g; i
k
NB
0fi:yk 0 0g; i
k k
(8)
where B is the cardinality of B , are typically introduced in Eq. (7) in order to exclude previous integer solutions. Pure binary constraints on the original system are also included in Eq. (7). The MIDO algorithm terminates when the difference between the least upper bound from the primal problems and the lower bound from the master problem is less than a specified tolerance, or if there is an infeasible master problem, in which case all feasible integer combinations have been considered. The solution to the MIDO problem then corresponds to the solution of the primal problem with the least upper bound.
where y is now a set of continuous, time-invariant search variables and y is the set of complicating (binary) variables. If the same values are chosen for the complicating variables in problems Eqs. (3) and (9), then these two problems are clearly identical. Thus, their k solutions are both denoted by uk (t) and d ; with resulting state variables xk (t) and algebraic variables d xk (t): The vector y at the optimum solution to Eq. (9) is a given by yk .
4.2. Construction of the master problem As before, the GBD relaxed master problem is formulated using dual information from the solution of the primal problem. It is obvious that the dual variables associated with hd, ha, h0, hp, gp, hq and gq are identical for the old and new primal formulations Eqs. (3) and (9). For a feasible primal problem Eq. (9), the Lagrange function, Lk ; is given by: k k k d a Lk (uk ; d ; y)0J(xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; yk ; tf ) '
4. An alternative approach In this section an alternative MIDO approach is proposed, still based on v2-GBD principles, but which does not require the computation of the dual variables via the complex integration in Eq. (4), for example, and which leads to the formulation of an equivalent but much more compact master problem than Eqs. (6) and
tf t0
T k k nk (t)hd (xd (t); xk (t); xk (t); uk (t); d ; yk ; t)dt d a d
652

tf t0
'
nk (t)ha (xk (t); xk (t); uk (t); d ; yk ; t)dt d a a
algorithm for the solution of the general MIDO problem Eq. (1) can be summarised as follows: 5.1. Step 1 Set the termination tolerance, o . Initialise the iteration counter, k 0/1; lower bound, LB 0/(/; and upper bound, UB0/. 5.2. Step 2 For fixed values of the binary variables, y0/yk , solve the k th primal problem Eq. (3) to obtain a solution Jk . (For computational speed, it may be useful to omit the search variables and constraints that are superfluous due to the current choice of values for the binary variables.) Set UB0/min(UB, Jk ) and store the values of the continuous and binary variables corresponding to the best primal solution so far. (10) 5.3. Step 3 Solve the primal problem Eq. (9) at the solution found in Step 2 with the full set of search variables and constraints included. Convergence will be achieved in one iteration. Obtain the multipliers vk needed to construct the k th master problem Eq. (12) 5.4. Step 4 Solve the k th relaxed master problem Eq. (12) with integer cuts Eq. (8) and the pure binary constraints of the original system, to obtain a solution hk . Update the lower bound, LB 0/hk . 5.5. Step 5 If UB(/LB 0/o , or the master problem is infeasible, then stop. The optimal solution corresponds to UB and the values stored in Step 2. Otherwise, set k 0/k'/1 and yk'1 equal to the integer solution of the k th master problem, and return to Step 2. Note that if in Step 2 a particular set of values yk renders the primal problem Eq. (3) infeasible, then an infeasibility minimisation problem is solved instead. This can involve, for example, minimising the l1 or l sum of constraint violations (Floudas, 1995). In Step 3, the infeasibility minimisation problem is then re-solved at its optimal solution to obtain multipliers vk . The formulation corresponding to minimising the l sum of violations is shown below:
a;u(t);d;y
T k k 'jk hd (xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; yk ; tf ) d a d;f
'jk ha (xk (tf ); xk (tf ); uk (tf ); d ; yk ; tf ) d a a;f 'jk hd (xd (t0 ); xk (t0 ); xk (t0 ); uk (t0 ); d ; yk ; t0 ) d a d;0
T k 'jk ha (xk (t0 ); xk (t0 ); uk (t0 ); d ; yk ; t0 ) d a a;0 T
'k
kT
k h0 (xk (t0 ); xk (t0 ); xk (t0 ); uk (t0 ); d ; yk ; t0 ) d d a
T k k 'mk hp (xd (ti ); xk (ti ); xk (ti ); uk (ti ); d ; yk ; ti ) d a p
T k k 'lk gp (xd (ti ); xk (ti ); xk (ti ); uk (ti ); d ; yk ; ti ) d a p
'mk hq (d ; yk )'lk gq (d ; yk ) q q 'vk (yk ( y);

T
where v is the set of dual multipliers for the equations hy 0/0. Since hd, ha, h0, hp and hq all equal 0 at the primal solution and they are not functions of the complicating variables y; their associated terms can be removed from the Lagrangian in Eq. (10). Similarly, the complementarity optimality conditions for the NLP in the primal problem ensure that lT gp 0lT gq 0 0; and since gp and p q gq are not functions of y either, their associated terms can also be removed. Hence, the Lagrange function Lk can be simplified to: k k k d a Lk (uk ; d ; y)0J(xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; yk ; tf ) 'vk (yk ( y); and the master problem posed as: min h;
y;h
T
(11)
k s:t: Lk (uk ; d ; y)0 h;
k Kfeas :
(12)
It is clear from the Lagrange function definition in Eq. (11) that the master problem Eq. (12) does not require any dual information with respect to the DAE system and so no intermediate adjoint problem is required for its construction. It should also be noted that the master problem Eq. (12) is exactly equivalent to the master problem Eq. (7). A proof of this can be found in Appendix A.
5. New MIDO Algorithm I Based on the theoretical developments presented so far, a new, generalised Benders decomposition-based
min a;
s:t: hd (xd (t); xd (t); xa (t); u(t); d; y; t)00; t [t0 ; tf ];
653
ha (xd (t); xa (t); u(t); d; y; t)00; t [t0 ; tf ]; h0 (xd (t0 ); xd (t0 ); xa (t0 ); u(t0 ); d; y; t0 )00; hp (xd (ti ); xd (ti ); xa (ti ); x(ti ); d; y; ti )00; i 01; . . . ; N; gp (xd (ti ); xd (ti ); xa (ti ); u(ti ); d; y; ti )0ae; i 01; . . . ; N; hq (d; y)00; gq (d; y)00; hy 0 y( yk 00; a E0; (13) ti [t0 ; tf ]; ti [t0 ; tf ];
k k d a s:t: J(xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; y; tf ) '
g g
tf t0 tf t0
T k k nk (t)hd (xd (t); xk (t); xk (t); uk (t); d ; y; t)dt d a d
'
nk (t)ha (xk (t); xk (t); uk (t); d ; y; t)dt d a a
T k k 'jk hd (xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; y; tf ) d a d;f
'jk ha (xk (tf ); xk (tf ); uk (tf ); d ; y; tf ) d a a;f 'jk hd (xd (t0 ); xk (t0 ); xk (t0 ); uk (t0 ); d ; y; t0 ) d a d;0 'jk ha (xk (t0 ); xk (t0 ); uk (t0 ); d ; y; t0 ) d a a;0 'kk h0 (xd (t0 ); xk (t0 ); xk (t0 ); uk (t0 ); d ; y; t0 ) d a
T k k 'mk hp (xd (ti ); xk (ti ); xk (ti ); uk (ti ); d ; y; ti ) d a p T T T
where a is a positive scalar quantity and e is a vector whose elements are all equal to one. In Step 4, the master problem Eq. (12) is then augmented as: min h;
y;h
T k k d a s:t: J(xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; yk ; tf )'vk (yk ( y)
0h; k Kfeas ; vk (yk ( y)00;

T
T k k 'lk gp (xd (ti ); xk (ti ); xk (ti ); uk (ti ); d ; y; ti ) d a p
'mk hq (d ; y)'lk gq (d ; y) q q k Kinfeas ; (14) 0h: (15)
where Kfeas denotes the set of iteration numbers for which feasible primal problems are found and Kinfeas the set for which infeasible primal problems are found.
6. Remarks on MIDO Algorithm I (1) As already stated, one advantage of the algorithm outlined above is that, even when the binary variables y appear explicitly (and non-linearly) in the DAE system, the master problem Eq. (12) does not require any dual information with respect to the DAE system and so no intermediate adjoint problem is required for its construction. This comes at the expense of adding more equations, hy 0/0, and search variables, y: However, these are only used in Step 3 when the primal problem is re-solved at its optimal solution, and convergence is achieved in just one iteration. The use of this novel strategy also circumvents any numerical difficulties (such as those suggested by Schweiger and Floudas, 1997) associated with reformulating the primal problem. (2) The master problem in Step 4 of the new MIDO algorithm has a much simpler form compared to algorithms where adjoint variables are required. In the latter case, the master problem defined by Eqs. (6) and (7) has the form: min h;
y;h
Construction of this master problem thus requires that the objective function and all the DAEs, equality and inequality constraints that involve the binary variables y be considered together with their adjoint and Lagrange multipliers. Their number may be considerably larger than the number of binary variables y, especially for complex models. In contrast, the master problem defined by Eqs. (11) and (12) for the new MIDO algorithm is: min h;
y;h
T k k d d s:t: J(xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; yk ; tf )'vk (yk ( y) 0h; (16)
which only requires the value of the primal objective function together with extra terms whose complexity is independent of the complexity of the process model. The practical benefit of this will be seen in Example 2, presented later in this article. (3) Another important benefit of the new MIDO algorithm is that it is independent of the type of method used for solving the dynamic optimization primal problem. The upshot of this is that with the formulation presented, any existing dynamic optimization solver can be used to solve MIDO problems since even the most elementary NLP codes give the simple Lagrange multi-
654
plier information required for the master problem. Note, however, that the advantages gained from the new algorithm do depend on the method used for the primal solution. The maximum benefits are obtained if control vector parameterisation is used since no evaluation of the adjoint variables is then necessary. If a multiple shooting approach is used, then the adjoints are already computed at the primal solution; however, the master problem formulated with Algorithm I is still much simpler and more attractive from a computational viewpoint than the master problem that would result otherwise. (4) Since the new MIDO algorithm is based on v2GBD, like the MIDO approaches of Mohideen et al. (1997) and Schweiger and Floudas (1997), it shares its limitations. In particular, the algorithm is only guaranteed to converge to the global optimum under convexity conditions for the primal problem with the binary variables appearing linearly and separably (Floudas, 1995). In the much more common, non-convex case, there is a possibility that the GBD relaxed master problems Eqs. (7) and (12) will exclude potentially feasible choices of y from the solution set. Note that for (MI)NLP problems, it has been shown that GBD can lead to solutions which are not even local optima if convexity criteria are not satisfied and a set of continuous variables is chosen as the complicating (fixed) variables for the primal problem (Floudas & Visweswaran, 1990; Sahinidis & Grossmann, 1991; Bagajewicz & Manousiouthakis, 1991). Perhaps in response to this, Barton, Allgor, Feehery and Galan (1998) and Allgor and Barton (1999) incorrectly suggested that the use of GBD approaches for MIDO has the property of convergence to an arbitrary point which is not even guaranteed to be a local minimum . In fact, the new MIDO algorithm (and the GBD algorithms of Mohideen et al., 1997 and Schweiger & Floudas, 1997) will always converge to at least a local optimum, where (Ross, Perkins, & Pistikopoulos, 1998): Definition. A local optimum of an MINLP is one which, for fixed binaries y, satisfies local optimality conditions for the resulting (primal) problem. The above definition is intuitively obvious since when y is discrete there can be no feasible neighbourhood around it which improves the objective. Therefore, the (local optimal) solution of any of the feasible primal problems will result in a local optimum of the MIDO problem.
extended to tackle MIDO problems which involve timedependent binary variables. Such problems can arise in areas such as batch scheduling and adaptive control, where in the latter, the control structure changes with time. The new MIDO algorithm presented in this article can also tackle problems involving time-dependent binary variables, which will be denoted by uy(t). The general formulation of such a problem is:
u(t);d;y;uy (t)
min
J(xd (tf ); xd (tf ); xa (tf ); u(tf ); d; y; uy (tf ); tf );
s:t: hd (xd (t); xd (t); xa (t); u(t); d; y; uy (t); t)00; t [t0 ; tf ]; ha (xd (t); xa (t); u(t); d; y; uy (t); t)00; t [t0 ; tf ]; h0 (xd (t0 ); xd (t0 ); xa (t0 ); u(t0 ); d; y; uy (t0 ); t0 )0 0; hp (xd (ti ); xd (ti ); xa (ti ); u(ti ); d; y; uy (ti ); ti )00; i 0 1; . . . ; N; gp (xd (ti ); xd (ti ); xa (ti ); u(ti ); d; y; uy (ti ); ti )00; i 0 1; . . . ; N; hq (d; y)00; gq (d; y)00: (17) ti [t0 ; tf ]; ti [t0 ; tf ];
uy(t) may be parameterised by means of a piecewiseconstant profile defined on Nuy sub-intervals over the time horizon [t0, tf]. Hence: uy (t)0uy (j); j 01; . . . ; Nuy : (18)
The values of uy(t) during the j th sub-interval are given by a set of time-invariant binaries uy(j). With these binaries fixed, Eq. (17) becomes a standard dynamic optimization problem. A similar approach to that described earlier can thus be employed. The primal problem in Step 3 of Algorithm I becomes:
u(t);d;y;[uy (j) j01;...;Nuy ]
min
J(xd (tf ); xd (tf ); xa (tf ); u(tf ); d; y; uy (tf ); tf );
s:t: hd (xd (t); xd (t); xa (t); u(t); d; y; uy (t); t)00; t [t0 ; tf ]; ha (xd (t); xa (t); u(t); d; y; uy (t); t)00; t [t0 ; tf ]; h0 (xd (t0 ); xd (t0 ); xa (t0 ); u(t0 ); d; y; uy (t0 ); t0 )00; hp (xd (ti ); xd (ti ); xa (ti ); u(ti ); d; y; uy (ti ); ti )00; i 01; . . . ; N; gp (xd (ti ); xd (ti ); xa (ti ); u(ti ); d; y; uy (ti ); ti )00; ti [t0 ; tf ]; ti [t0 ; tf ];
7. Extending Algorithm I to handle time-dependent binary variables Ross et al. (1998) showed that the GBD-based MIDO algorithm of Mohideen et al. (1997) can be readily
i 01; . . . ; N; hq (d; y)00; gq (d; y)00;
655
hy 0y( yk 00; huy (j)0uy (j)( uk (j)00; y j 01; . . . ; Nuy : (19)
s:t: h EJ ' '
@J
In Eq. (19) uy(j ), j0/1,. . ., Nuy , and y are continuous search variables, while uy (j); j 0 1; . . . ; Nuy ; and y are the complicating, binary variables. The constraints huy (j)00; j 01; . . . ; Nuy ; are interior-point constraints that are handled in a similar fashion to the point constraints hp 0/0. The multipliers associated with huy(j) are denoted by vuy(j). The values of the variables at the optimum are denoted as before with the addition of uk (j): The master problem is formulated in a similar y manner to Eq. (14);
y;[uy ; j01;...;Nuy ];h
@d k X @J
j
k (d( d )' (uj ( uk ); j

k
@J @y
(y(yk )
k
@uj
@hp @hp k 0 ETk hk ' (d( d )' (y(yk ) p p @d k @y k X@h p ' (uj ( uk ) j @uj k j
j j
0 Egk ' p
min
h;
Nuy X j01
@gp k )' @gp (y(yk ) (d( d @d k @y k X@gp ' (uj ( uk ); j @uj k j
J 'v (y ( y)' k Kfeas ;
kT
[vky (j)]T [uk (j)( uy (j)]0h; y u
@hq @hp k 0 ETk hk ' (d( d )' (y(yk ) ; p q @d k @y k
Nuy X [vky (j)]T [uk (j)( uy (j)] 00; y v (y ( y)' u kT k j01
k Kinfeas : (20)
0 Egk ' q
@gq @hp k (d( d )' (y(yk ); @d k @y k
0 0y( y; k Kfeas ; (21)
8. Using outer-approximation to construct the master problem As discussed in the review section earlier in this article, current MIDO algorithms that use the principles of OA/ER to construct the master problem are limited in that they are not applicable to models where the binary variables appear explicitly in the DAE system. This can be overcome using the reformulation strategy that was applied in Algorithm I. In an OA/ER-based MIDO algorithm, the primal problems are solved in an identical manner to those in Step 2 and Step 3 of the GBD-based MIDO Algorithm I. However, in order to construct the master problems, instead of obtaining Lagrange multiplier values in Step 3, gradient information is required of the objective function and constraints with respect to the optimization search variables. During solution of the primal problem, the independent search variables are d, y and the time-invariant variables resulting from the parameterisation of the controls, which are denoted by uj. The k th master problem (for the case of feasible primal problems) is thus:
y;d;y;[uj ;j ];h
together with the integer cuts Eq. (8) and pure binary constraints of the original system. In Eq. (21), Tk , 0/ p, q, are diagonal matrices with elements defined as: Tk 0f(1 if lk B0; ;i ;ii '1 if lk 0; ;i 0 if lk 0 0; ;i where lk , 0/p, q are the Lagrange multipliers ,i pertaining to the equality constraints hp 0/0 and hq 0/ 0. If control vector parameterisation is used to solve the primal problem, then these multipliers and the gradients in Eq. (21) are available directly from the solution of the resulting NLP problem. If another type of dynamic optimization technique is used, e.g. multiple shooting, then the multipliers will still be directly available; however, the gradients have to be computed by performing an additional integration of the sensitivity or adjoint-based DAE system at the optimal solution of the primal problem (Rosen & Luus, 1991). Note that Eq. (21) bears some resemblance to the master problem proposed by Fletcher and Leyffer (1994). However, that work considers MINLP problems, not MIDO ones, where there are no equality constraints and where linearisations of inactive constraints are not included in the master problem.
min
h;
656
9. New MIDO Algorithm II A new algorithm, based on OA/ER, for the solution of the general MIDO problem Eq. (1) can now be summarised as follows: 9.1. Step 1 Set the termination tolerance, o . Initialise the iteration counter, k 0/1; lower bound, LB 0/(/; and upper bound, UB 0/. 9.2. Step 2 For fixed values of the binary variables, y 0/yk , solve the k th primal problem Eq. (3) to obtain a solution Jk. (For computational speed, it may be useful to omit the search variables and constraints that are superfluous due to the current choice of values for the binary variables.) Set UB 0/min(UB, Jk ) and store the values of the continuous and binary variables corresponding to the best primal solution so far. 9.3. Step 3 Solve the primal problem Eq. (9) at the solution found in Step 2 with the full set of search variables and constraints included. Convergence will be achieved in one iteration. Obtain the gradients and Lagrange multipliers needed to construct the k th master problem Eq. (21). 9.4. Step 4 Solve the k th relaxed master problem Eq. (21), with integer cuts Eq. (8) and the pure binary constraints of the original system, to obtain a solution hk . Update the lower bound, LB 0/hk . 9.5. Step 5 If UB(/LB 0/o , or the master problem is infeasible, then stop. The optimal solution corresponds to the values stored in Step 2. Otherwise, set k 0/k'/1 and yk'1 equal to the integer solution of the k th master problem, and return to Step 2. As with the GBD-based MIDO Algorithm I, if in Step 2 a particular set of values yk renders the primal problem infeasible, then an infeasibility minimisation problem can be solved, which in Step 3 may correspond to Eq. (13) (Viswanathan & Grossmann, 1990). The matrices Tp and Tq and gradients @ hp/@ d, @ gp/@ d, @ hq/ @ d, @ gq/@ d etc. are found from the solution of this problem and then used to add extra linearisation constraints to the master problem Eq. (21).
10. Remarks on MIDO Algorithm II (1) As already stated, unlike current MIDO approaches that use OA principles to construct the master problems, the algorithm outlined above is applicable to models where the binary variables appear explicitly in the DAE system. Like Algorithm I, Algorithm II has the advantage that it is independent of the type of method that is used for solving the dynamic optimization primal problem (although control vector parameterisation may be preferred versus multiple shooting, for example, since then no intermediate problem needs to be solved to obtain the gradients that are necessary for constructing the master problem). Thus, again as with Algorithm I, Algorithm II allows existing dynamic optimizers to be used for solving MIDO problems. Algorithm II can also be readily extended to tackle problems involving timedependent binary variables, in a similar fashion to that described in Section 7 for Algorithm I. (2) It should be noted that unlike the GBD-based MIDO Algorithm I, where the master problem Eq. (16) involves just one additional constraint per iteration, the master problem Eq. (21) in Algorithm II involves the addition of a large number of constraints at each iteration. These constraints correspond to linearisations about the optimal solution of the primal problem of the objective function, equality and inequality constraints. The MILP master problems in Algorithm II are therefore computationally more expensive than those in Algorithm I. (As noted in Section 2 of this paper, however, for both GBD- and OA-based MIDO algorithms, the master problems are usually much less computationally demanding to solve than the primal, dynamic optimization problems.) Because the master problem in Algorithm II is more constrained than that in Algorithm I, the lower bound will, in general, be tighter, and so Algorithm II will often converge to a locally optimal solution of the MIDO problem in fewer iterations than Algorithm I. Note, however, that since most MIDO problems involve non-convexities, the additional constraints in the master problem of Algorithm II may cause it to cut off larger parts of the feasible region than Algorithm I. Thus, although Algorithm II may converge to a local optimum faster than Algorithm I, there is a chance that the solution will be worse.
11. Example 1 The application of the new algorithms for solving MIDO problems is now demonstrated with an example based on a binary distillation system from the MINOPT Users Guide (Schweiger, Rojnuckarin & Floudas, 1997). The column has a fixed number of trays and the objective is to determine the optimal feed location
657
(discrete decision), vapour boil-up, V , and reflux flow rate, R (continuous decisions), in order to minimise the integral square error (ISE) between the bottoms and distillate compositions and their respective set-points. The superstructure of the system is shown in Fig. 1. The following modelling assumptions were used by Schweiger et al.: (i) constant molar overflow; (ii) constant relative volatility, a ; (iii) phase equilibrium; (iv) constant liquid hold-ups, equal to m for each tray and 10m for the reboiler and condenser; (v) no tray hydraulics; (vi) negligible vapour hold-ups; and (vii) no pressure drops. The system is initially at steady-state; at t0/0 there is a step change in the feed composition, zf; and the inequality constraints are that the distillate composition must be greater than 0.98, and the bottoms composition must be less than 0.02, at the end of the time horizon of 400 min. The problem can be stated mathematically as: (1) Objective:
fyf1 ;...;yfN ;V ;Rg
10m
dxN'1 dt
0V (yN (xN'1 );
(3) Overall balances: 0 0L1 (V (B; 0 0Li'1 (Li 'Fyfi ; i 01; . . . ; N (1;
0 0(LN 'V 'FyfN 'R; 0 0V (D(R; (4) Vapour /liquid equilibrium: y0 0 axb ; 1 ' (a ( 1)xb axi 1 ' (a ( 1)xi ; i 01; . . . ; N;
yi 0
min
ISE(tf );
(5) Initial conditions: ISE(t0 0)0 0;
d(ISE) dt
0 (xb (x+ )2 '(xN'1 (x+ )2 ; b N'1
dxb dt dxi dt
s.t. (2) Component balances: dx 10m b 0L1 x1 (Vy0 (Bxb ; dt m dxi dt 0Li'1 xi'1 (Li xi 'V (yi(1 (yi )'F yfi zf ;
j j
00;
t00
0 0;
i 01;
; N '1;
t00
(6) Step disturbance (smoothly approximated): zf 00:54(0:09e(10t ; (7) End-point inequality constraints: g1 00:98(xN'1 (tf )00; g2 0xb (tf )(0:020 0; (8) Time-horizon: t [0; 400]; (9) Pure binary constraints:
N X i01
i 01; . . . ; N (1; m dxN 0(LN xN 'V (yN(1 (yN )'F yfN zf 'RxN'1 ; dt
yfi 01; i 01; . . . ; 9; i 010; . . . ; N;
yfi 0 0;
yfi f0; 1g;
Fig. 1. Superstructure for Example 1.
where yfi is a binary variable that takes a value of 1 if tray i receives the feed, and is 0 otherwise. Only one tray is allowed to receive the feed, which is constrained to enter on tray 10 or above. The MIDO Algorithms I and II were both implemented using gPROMS/gOPT v1.7 to solve the dynamic optimization primal problems, and GAMS/CPLEX v2.50 (Brooke, Kendrick & Meeraus, 1992) for the MILP master problems. All the primal solutions,
658
V. Bansal et al. / Computers and Chemical Engineering 27 (2003) 647 /668 Table 1 Progress of iterations of MIDO algorithm I for example 1 Iteration number Primal solutions Feed tray V (kmol/min) R (kmol/min) ISE UB Lagrange multipliers v1 v2 v3 v4 v5 v6 v7 v8 v9 v10 v11 v12 v13 v14 v15 v16 v17 v18 v19 v20 v21 v22 v23 v24 v25 v26 v27 v28 v29 v30 Master solutions Feed tray LB UB(/LB0/0? 1 2 3 4
Lagrange multipliers and master solutions using Algorithm I are summarised in Table 1, while the primal solutions, gradients and master solutions using Algorithm II are summarised in Table 2. Using a termination tolerance of o0/0, both algorithms converged after four iterations, giving the same optimal solution reported by Schweiger et al. (1997). This corresponds to a feed location at tray 25, vapour boil-up of 1.5426 kmol/min, reflux flow rate of 1.0024 kmol/min, and ISE of 0.1817. Note that integer cuts were added in each master problem to prevent the re-occurrence of previous solutions. For example, for the first master problem, the cut (yf20 (ai"20 yfi 00); was added. It is interesting to observe that, for this example, the master problems of Algorithm II do not predict tighter lower bounds than the master problems of Algorithm I except on the fourth iteration when convergence of the overall MIDO problem has already been achieved. This behaviour can be understood in terms of the fact that, qualitatively, the difference between a GBD-based master problem and an OA-based one is that the former only considers the inequality constraints that are active at the primal solution, whereas the latter considers all the inequality constraints in the model (Duran, 1984). Thus, if the inequality constraints that are inactive at the primal solution remain inactive at the solution of the OA master problem, then GBD and OA will give identical behaviour. This is demonstrated for this example in Appendix B.
20 1.2231 0.6836 0.1992 0.1992 0.4353 0.3319 0.2741 0.2414 0.2226 0.2114 0.2045 0.2001 0.1970 0.1948 0.1931 0.1916 0.1904 0.1893 0.1884 0.1878 0.1876 0.1879 0.1890 0.1909 0.1994 0.2085 0.2172 0.2246 0.2301 0.2325 0.2298 0.2189 0.1954 0.1535 26 0.1576 No
26 1.7830 1.2426 0.1827 0.1827 0.6147 0.3965 0.2864 0.2307 0.2024 0.1881 0.1808 0.1771 0.1753 0.1745 0.1742 0.1742 0.1744 0.1746 0.1749 0.1752 0.1755 0.1759 0.1763 0.1766 0.1770 0.1773 0.1777 0.1779 0.1777 0.1765 0.1698 0.1537 0.1204 0.0566 24 0.1813 No
24 1.4111 0.8711 0.1834 0.1827 0.6045 0.4077 0.3027 0.2463 0.2159 0.1992 0.1899 0.1847 0.1815 0.1796 0.1784 0.1775 0.1768 0.1763 0.1759 0.1755 0.1752 0.1749 0.1747 0.1746 0.1748 0.1753 0.1762 0.1776 0.1818 0.1846 0.1838 0.1763 0.1574 0.1209 25 0.1815 No
25 1.5426 1.0024 0.1817 0.1817 0.6204 0.4087 0.2984 0.2406 0.2101 0.1939 0.1853 0.1806 0.1780 0.1765 0.1757 0.1752 0.1749 0.1747 0.1746 0.1745 0.1745 0.1745 0.1745 0.1746 0.1747 0.1750 0.1754 0.1759 0.1763 0.1763 0.1721 0.1600 0.1339 0.0850 23 0.1848 Yes: STOP
12. Example 2 This example, adapted from Schweiger and Floudas (1997), builds on Example 1, where the distillation system is now subject to sinusoidal disturbances in the feed flow-rate (9/20% variation, 3-h period) and composition (9/20% variation, 1-h period). The objective is to design the distillation column and its required control system at minimum total annualised cost (comprising capital cost of the column and operating cost of utilities), capable of feasible operation over the whole of a given time horizon (1440 min). Feasibility is defined through the satisfaction of inequality path constraints related to the product quality specifications and the minimum allowable column diameter required to avoid flooding. It is assumed that proportional-integral (PI) controllers are to be used, where the vapour boil-up rate, V , controls the bottoms composition, xb, and the reflux rate, R , controls the distillate composition, xN'1. Solution of the problem thus requires the determination of: (i) the optimal process design, in terms of the number of trays (discrete decision), feed tray location, which is constrained to lie at least five trays below the reflux tray (discrete decision), and column diameter, Dcol (continuous decision); and (ii) the optimal control design, in
terms of the gains (Kv and Kr ), reset times (tv and tr ) and set-points (x1 and x1 ) of the two PI controllers. b N'1 The superstructure of this system, together with relevant parameter values and definitions of variables, is shown in Fig. 2. Using similar modelling assumptions to those in Example 1, the MIDO problem can be stated mathematically as: (1) Objective: min Cost
fyf ;yr;Dcol ;Kv ;Kr ;tv tr ;x+ ;x+ g b N'1
07756Vmean '1891:125'996:3D2 col
V. Bansal et al. / Computers and Chemical Engineering 27 (2003) 647 /668 Table 2 Progress of iterations of MIDO algorithm II for example 1 Iteration number Primal solutions Feed tray V (kmol/min) R (kmol/min) ISE UB Objective gradients w.r.t. V , R yf1, yf2 yf3, yf4 yf5, yf6 yf7, yf8 yf9, yf10 yf11, yf12 yf13, yf14 yf15, yf16 yf17, yf18 yf19, yf20 yf21, yf22 yf23, yf24 yf25, yf26 yf27, yf28 yf29, yf30 Inequalities values g1, g2 Gradients of g1 w.r.t. V , R yf1, yf2 ()/1e6) yf3, yf4 ()/1e6) yf5, yf6 ()/1e6) yf7, yf8 ()/1e6) yf9, yf10 ( )/1e6) yf11, yf12 ( )/1e7) yf13, yf14 ( )/1e7) yf15, yf16 ( )/1e6) yf17, yf18 ( )/1e5) yf19, yf20 ( )/1e5) yf21, yf22 ( )/1e4) yf23, yf24 ( )/1e3) yf25, yf26 ( )/1e3) yf27, yf28 ( )/1e2) yf29, yf30 Gradients of g2 w.r.t. V , R yf1, yf2 yf3, yf4 yf5, yf6 yf7, yf8 yf9, yf10 yf11, yf12 yf13, yf14 yf15, yf16 yf17, yf18 yf19, yf20 yf21, yf22 yf23, yf24 yf25, yf26 yf27, yf28 yf29, yf30 Master solutions 1 2 3 4
659
20 1.2231 0.6836 0.1992 0.1992 7.4445, (/6.7611 (/4.9048, (/4.8013 (/4.7434, (/4.7107 (/4.6918, (/4.6806 (/4.6737, (/4.6692 (/4.6662, (/4.6639 (/4.6622, (/4.6608 (/4.6596, (/4.6587 (/4.6581, (/4.6580 (/4.6588, (/4.6611 (/4.6657, (/4.6743 (/4.7129, (/4.7640 (/4.8258, (/4.8989 (/4.9855, (/5.0899 (/5.2187, (/5.3816 (/5.5935, (/5.8772 (/3.9013e(/3, 0 0.1819, (/0.3292 (/385.63, (/415.24 (/433.12, (/444.65 (/452.71, (/458.81 (/463.70, (/467.65 (/470.56, (/471.93 (/4706.6, (/4645.9 (/4497.0, (/4185.2 (/357.15, (/239.88 (/1.9069, 39.354 116.14, 258.75 91.024, 181.41 29.602, 43.746 61.235, 83.232 11.157, 1.4905 0.1998, 0.2700 (/1.8800, 1.7074 1.1287, 1.1287 1.1287, 1.1287 1.1287, 1.1286 1.1286, 1.1286 1.1286, 1.1286 1.1286, 1.1286 1.1287, 1.1287 1.1288, 1.1289 1.1292, 1.1296 1.1305, 1.1322 1.1398, 1.1504 1.1639, 1.1804 1.2009, 1.2267 1.2599, 1.3038 1.3632, 1.4455
26 1.7830 1.2426 0.1827 0.1827 12.108, (/11.926 (/7.3116, (/7.0977 (/6.9898, (/6.9352 (/6.9075, (/6.8934 (/6.8863, (/6.8827 (/6.8809, (/6.8802 (/6.8799, (/6.8798 (/6.8800, (/6.8802 (/6.8805, (/6.8808 (/6.8812, (/6.8817 (/6.8823, (/6.8831 (/6.8845, (/6.8869 (/6.8913, (/6.8998 (/6.9163, (/6.9487 (/7.0640, (/7.2642 (/7.6064, (/8.1968 (/2.2065e(/3, 0 0.0423, (/0.0684 (/1.0724, (/1.1254 (/1.1541, (/1.1703 (/1.1796, (/1.1838 (/1.1825, (/1.1722 (/1.1448, (/1.0837 (/9.5487, (/6.9007 (/1.5164, 9.3740 3.1344, 7.5611 1.6475, 3.4417 7.0530, 14.321 2.8946, 5.8379 1.1761, 2.3680 4.7665, 9.5933 2.7123, 5.8186 0.1121, 0.2061 (/2.0744, 2.0432 1.1473, 1.1481 1.1484, 1.1486 1.1487, 1.1488 1.1488, 1.1488 1.1488, 1.1488 1.1488, 1.1488 1.1488, 1.1488 1.1488, 1.1488 1.1488, 1.1489 1.1489, 1.1489 1.1491, 1.1495 1.1502, 1.1516 1.1544, 1.1602 1.1811, 1.2182 1.2825, 1.3946
24 1.4111 0.8711 0.1834 0.1827 10.052, (/9.7077 (/6.2937, (/6.0972 (/5.9924, (/5.9361 (/5.9057, (/5.8891 (/5.8798, (/5.8746 (/5.8715, (/5.8695 (/5.8683, (/5.8674 (/5.8668, (/5.8663 (/5.8659, (/5.8656 (/5.8655, (/5.8655 (/5.8660, (/5.8673 (/5.8700, (/5.8754 (/5.8856, (/5.9049 (/5.9830, (/6.1007 (/6.2709, (/6.5180 (/6.8835, (/7.4364 (/2.9734e(/3, 0 0.0821, (/0.1403 (/11.278, (/12.010 (/12.433, (/12.692 (/12.862, (/12.976 (/13.045, (/13.058 (/12.981, (/12.740 (/121.89, (/110.47 (/87.721, (/43.272 4.2762, 20.850 5.2700, 11.383 23.107, 45.587 8.8680, 17.128 3.2961, 6.3307 18.824, 38.274 6.722, 11.031 0.1754, 0.2751 (/1.9993, 1.9309 1.1316, 1.1317 1.1317, 1.1317 1.1317, 1.1317 1.1317, 1.1317 1.1317, 1.1317 1.1317, 1.1317 1.1317, 1.1317 1.1317, 1.1318 1.1318, 1.1319 1.1320, 1.1323 1.1328, 1.1338 1.1356, 1.1392 1.1539, 1.1767 1.2107, 1.2614 1.3378, 1.4551
25 1.5426 1.0024 0.1817 0.1817 10.962, (/10.701 (/6.7519, (/6.5413 (/6.4315, (/6.3740 (/6.3438, (/6.3277 (/6.3191, (/6.3144 (/6.3118, (/6.3103 (/6.3095, (/6.3090 (/6.3087, (/6.3085 (/6.3084, (/6.3084 (/6.3085, (/6.3087 (/6.3092, (/6.3101 (/6.3119, (/6.3153 (/6.3220, (/6.3349 (/6.3594, (/6.4534 (/6.6042, (/6.8391 (/7.2083, (/7.7990 (/2.6192e(/3, 0 0.0615, (/0.1024 (/3.7995, (/4.0215 (/4.1467, (/4.2212 (/4.2676, (/4.2958 (/4.3067, (/4.2933 (/4.2375, (/4.1012 (/38.088, (/32.120 (/20.205, 3.3290 4.9574, 14.021 3.1759, 6.6453 13.429, 26.689 5.2608, 10.327 2.0229, 3.9582 7.7409, 22.509 4.6826, 8.5604 0.1477, 0.2481 (/2.0286, 1.9803 1.1347, 1.1349 1.1350, 1.1351 1.1350, 1.1351 1.1350, 1.1351 1.1350, 1.1351 1.1351, 1.1351 1.1351, 1.1351 1.1351, 1.1351 1.1351, 1.1352 1.1353, 1.1354 1.1357, 1.1363 1.1375, 1.1398 1.1442, 1.1617 1.1903, 1.2360 1.3092, 1.4276
660 Table 2 (Continued ) Iteration number Feed tray V , R (kmol/min) LB UB(/LB0/0? 1
2 24 1.5179, 0.9777 0.1813 No
3 25 1.6375, 1.0977 0.1815 No
4 23 1.3856, 0.8440 0.2093 Yes: STOP
26 1.7488, 1.2072 0.1576 No
N X '1494:45Dcol 6'0:76 i yri

i01
10m
dxN'1 0 V (yN (xN'1 ); dt 0:0014134 Dcol 2

3
'61:25
N X i01
i yri (0:7'1:5D2 ); col
m 07:538115D2 col
'hweir ;
(3) Overall balances: dVmean V 0 ; tf dt s.t. (2) Component balances: 10m X N

i0 0i
0 0L1 (V (B; 0 0Li'1 (Li 'Fi 'Ri ; 0 0V (D(R; (4) Vapour /liquid equilibrium: y0 0 axb 1 ' (a ( 1)xb axi 1 ' (a ( 1)xi ; ; i 0 1; . . . ; N;
dxb dt
0 L1 x1 (Vy0 (Bxb ;
yri? m
dxi dt
0Li'1 xi'1 (Li xi 'V (yi(1 (yi )
yi 0
i 01; . . . ; N;
'Fi zf 'Ri xN'1 ; Fi 0F yfi ; Ri 0R yri ; LN'1 0 0;
i 0 1; . . . ; N;
(5) Bottoms composition control loop: V 0Vss 'Kv (ev 'Iv ); ev 0 x+ (xmeas ; b b
i 01; . . . ; N; i 0 1; . . . ; N;
dIv ev 0 ; dt tv dVss 0 0; dt (6) Distillate composition control loop: R 0Rss 'Kr (er 'Ir ); er 0 x+ (xmeas ; N'1 N'1 dIr dt 0 er tr ;
dRss dt
0 0;
(7) Measurement lags:

Fig. 2. Superstructure for Example 2.
xb 01:2
dxmeas b dt
'xmeas ; b
661
xN'1 01:2
dxmeas N'1 dt
'xmeas ; N'1
yfi (
N X i?0i'5
yri? 0 0;
i 01; . . . ; N;
(8) Initial conditions: Vmean (t00)00; dxb dt dxi dt
yfi ; yri f0; 1g;
i 01; . . . ; N;
00;
t00
00;
i 0 1; . . . ; N '1;
t00
ev (t00)00; Iv (t0 0)0 0; er (t00)00; Ir (t0 0)00; dxmeas b dt
0 0;
t00
dxmeas N'1 dt
00;
t00
(9) Sinusoidal disturbances: 2pt ; F 01'0:2sin 180 zf 00:45'0:09sin 2pt 60 ;
(10) Inequality path constraints: g1 0 0:98(xN'1 (tf )00; g2 0 xb (tf )(0:0200; g3 0 Dmin (Dcol 0 0; col p Dmin 00:6719 V ; col (11) Time horizon: t [0; 1440]; (12) Pure integer constraints:
N X i01 N X i01
yfi 01;
yri 01;
where the new variables that have been defined compared to Example 1 are Vmean, the average boil-up rate over the time horizon; Dmin, the minimum allowable col column diameter due to flooding; and yri, a binary variable that takes a value of 1 if tray i receives the reflux, and is 0 otherwise. Note that the model above is similar to that used by Schweiger and Floudas (1997), although in that study the simpler case was considered of a single step disturbance and only end-point constraints instead of path constraints. It should also be noted that the binary variable term, (aN yri? ); in the i?01 component balances for the trays, is needed in order to ensure that no separation is effected in the shadow trays above the tray that receives the reflux. Without this term, the vapour flow leaving the reflux tray would be allowed to come into contact with the liquid hold-up on the shadow trays. Omission of the term would thus cause the properties of the vapour outlet from the top tray N to be different from those of the vapour outlet from the reflux tray, rendering such a superstructure model invalid for determining the optimal number of trays. After conversion of the three path constraints into end-point inequality constraints using the formulation described by Bansal (2000), the MIDO problem was solved using Algorithm I, again implemented within gPROMS/gOPT v1.7 and GAMS v2.50/CPLEX. Despite the highly combinatorial nature of the problem (465 discrete alternatives), the algorithm converged in just four primal iterations. The primal and master design and control solutions at each iteration are shown in detail in Table 3. The optimal solution corresponds to a distillation column with 15 trays and the feed located on tray 8. The profiles of the controlled distillate and bottoms compositions for the optimal distillation system are shown in Figs. 3 and 4, respectively. It can be seen that the path constraints g1 0/0.98(/xN'1 0/0 and g2 0/ xb(/0.02 0/0 are satisfied at all times over one weeks operation, even though the MIDO problem was only solved using a time horizon of one day. Fig. 5 shows the profile of the minimum allowable column diameter due to flooding and compares it with the optimal column diameter. Again, the associated path constraint, g3 0 Dmin (Dcol 00; is always satisfied. The cyclical behacol viour observed in Figs. 3 /5 and the fact that all three path constraints are satisfied at all times suggests that the solution found is not only economically optimal but is also stable for the set of disturbances considered. Finally, as stated in Remark 2 of Section 6, earlier in this article, one advantage of using Algorithm I compared to current MIDO algorithms such as those of
662

35 X i01 35 X i01
Table 3 Progress of iterations of MIDO algorithm I for example 2 Number Primal solutions Discrete decisions No. of trays Feed tray Process design Dcol (m ) Bottoms control loop Kv tv 1 xb Distillate control loop Kr tr 1 xN'1 Cost ($k yr (1) UB Master solutions No. of trays Feed tray LB UB(/LB0/0 1 2 3 4
s:t: h ECostp ' k Kfeas ;
wfik (yfik (yf i )'
wrk (yrk (yri ); i i
(22)
20 7 0.8214 (/55.92 5.66 0.01823 44.90 654.2 0.9836 40.303 40.303 14 9 32.989 No 14 9 0.9138 (/43.32 727.1 0.01476 58.70 27.2 0.9824 39.633 39.633 15 8 34.502 No 15 8 0.8458 (/54.64 155.4 0.01564 23.20 2.01 0.9806 36.729 36.729 16 8 36.349 No 16 8 0.8323 (/43.83 7.00 0.01747 50.40 211.5 0.9830 36.901 36.729 15 7 36.879 Yes: STOP
where the pure binary constraints have not been included for clarity of presentation; wfi are the Lagrange k multipliers for the constraints yfi (yf i 0 0; i 0 1; . . . ; 35; that are added when applying Step 3 of Algorithm I; and wri are the Lagrange multipliers for the constraints yri (yrk 00; i 01; . . . ; 35: These muli tipliers are all directly reported by the optimizer used for solving the primal problems. In sharp contrast, the master problem if Algorithm I is not used is:
yfi(1;...;35) ;yri(1;...;35)
min
h;
p h E7756Vmean '1891:125'996:3(Dk )2 col
35 X '1494:45Dcol 6'0:76 i yri

i01 35 X i01 35 1440 X 0 i01
Mohideen et al. (1997) and Schweiger and Floudas (1997), is the far simpler form of the master problem. This can be seen very clearly for this example. Using Algorithm I, the k th master problem for this example has the form:
'61:25
i yri [0:7'1:5(Dk )2 ] col X 35 dxk yri? m i (Lk xk 'Lk xk i'1 i'1 i i dt i?0i
'
g g
nk 1;i
min
yf i(1;...;35) ;yri(1;...;35)
h;
(V k (yk (yk )(Fi zf (Rk xk ]gdt i(1 i i N'1 '

35 1440 X 0 i01
[nk (Fi (Fyfi )]dt; 2;i
Fig. 3. Distillate composition prole for the optimal solution of Example 2.
663
Fig. 4. Bottoms composition prole for the optimal solution of Example 2.
'
35 1440 X 0 i01
[nk (Rk (Rk yri )]dt; 3;i i
13. Concluding remarks k Kfeas ; (23) This article has presented a new algorithm for solving MIDO problems that is based on generalised Benders decomposition principles with the use of a reformulation strategy. This algorithm has many advantages over current approaches, namely that (i) it is independent of the type of method used to solve the dynamic optimization primal problems; (ii) no intermediate adjoint problem needs to be solved in order to construct the master problems, even when the binary variables appear explicitly and non-linearly in the DAEs; and (iii) the master problems are significantly simpler. The algorithm can also be easily extended to handle the case of time-dependent binary variables, as shown in the article.
where n1,i , n2,i and n3,i are the time-dependent dual (adjoint) multipliers for the DAEs that contain the binary variables, the values of which can only be obtained by solving a complex, intermediate problem. It is clear that Eq. (23) is much more complicated than Eq. (22), since it involves integral terms and the storage of a large number of primal solution values (e.g. the values of dxk =dt; Lk , xk , yk etc.). i'1 i'1 i(1 i
Fig. 5. Minimum column diameter prole for the optimal solution of Example 2.
664
With the reformulation strategy used in Algorithm I, the article has also demonstrated how an algorithm can be developed that uses OA/ER principles to construct the master problems. This algorithm has similar properties to Algorithm I and thus, in contrast to current OAbased, MIDO algorithms, it is applicable to models where the binary variables appear explicitly in the DAEs. The algorithms presented in this article are well-suited for the practical solution of large engineering problems using available tools. This offers the potential for much progress to be made in the solution of more realistic problems in areas such as simultaneous design and control under uncertainty (Bansal, Pistikopoulos & Perkins, 2002; Bansal, Ross, Pistikopoulos & Perkins, 2002).
Note that in Eq. (24), all of the Jacobian matrices appear in transposed form. Thus, (@ J/ @y) is a (1 )/ny) vector, (@ hd/ @y) is a (nxd )/ny) matrix, and so on, where ny and nxd are the numbers of binary and differential variables, respectively. Substituting the expression for vk given by Eq. (24) into Eq. (11) then gives: k k k d a Lk (uk ; d ; y)0J(xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; yk ; tf ) @J ' ' @y yk
T 'jk d;f
@hd @y
j g j

tf t0
T nk d
@hd @y
dt'
yk
tf t0
T nk a
@ha dt @y yk
j j
T 'jk a;f
@ha @y
tf ;yk
tf ;yk
T 'jk d;0
@hd @y
T 'jk a;0
@ha @y
Acknowledgements The authors gratefully acknowledge nancial support from the EPSRC, the Centre for Process Systems Engineering Industrial Consortium, DETR/ETSU and Process Systems Enterprise Ltd. V. Bansal would also like to thank ICI (Ph.D. Scholarship) and the Society of the Chemical Industry (SCI Messel Scholarship). 'mk p
T
t0 ;yk
'k
t0 ;yk
kT
@h0 @y
yk
@hp @gp T 'lk p @y yk @y yk
@hq @gq T T 'mk 'lk (y(yk ): q q @y yk @y yk
(25)
Appendix A: Proof that master problems Eq. (12) and Eq. (7) are equivalent In order to show that master problems Eq. (12) and Eq. (7) are equivalent, equivalence needs to be estab lished between the Lagrange functions Lk in Eq. (11) k and L in Eq. (6). This can be achieved as follows. Firstly, consider the optimality conditions that must be satisfied by the continuous variables y in Eq. (9). These are determined by finding a stationary point for the Lagrange function Lk in Eq. (10), i.e.: T T tf tf @J @hd @ha T k 00 ' nd dt' nk dt a @y @y @y t0 t0 k k k y y y
Now consider the original Lagrange function Lk , Eq. (6), used in the master problems of the algorithms of Mohideen et al. (1997) and Schweiger and Floudas (1997). For a system that is linear in y (as those algorithms require), the objective function term in Eq. (6) can be decomposed as: k k J(xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; y; tf ) d a 0J1 (xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; tf ) d a 'J2 (xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; tf )y: d a k k J(xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; yk ; tf ) d a 0J1 (xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; tf ) d a 'J2 (xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; tf )yk : d a Eliminating J1 from Eq. (26) and Eq. (27) gives: k k J(xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; y; tf ) d a 0J(xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; yk ; tf ) d a 'J2 (xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; tf )(y(yk ): d a
k k k k k k k k k k k k
(26)
Similarly, the objective function evaluated at y0/yk is:
j g
T @hd ' @y T @hd ' @y
j j j
jk ' d;f
tf ;yk
T @ha @y @ha @y
jk a;f T @h0 @y
(27)
tf ;yk
jk ' d;0
t0 ;yk
T
jk ' a;0
t0 ;yk
kk
yk
T T T @hp @gp @hq k k ' mp ' lp ' @y @y @y yk yk T @gq ' lk 'Ivk : q @y yk
mk q (24)
(28)
yk
Now, J2 is simply equal to (@J/@y) and so Eq. (28) can be written as: k k J(xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; y; tf ) d d
665
k k 0 J(xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; yk ; tf )' d a (y(y ):

k
@J @y
yk
Appendix B: Demonstrating the equivalence of GBD and OA for Example 1 (29) In Step 3 of Algorithm I/II, a dynamic optimization problem is solved. During the course of this, the optimizer sees a NLP formed from the elimination of the differential and algebraic variables during an integration phase. For Example 1, this NLP takes the form:
V ;R;yfi(i01;...;30)
In an analogous manner to that above, the terms within the Lagrange function Eq. (6) that relate to the equality constraints can be written as: @h h (. . . ; y . . .)0 h (. . . ; yk . . .)' (y(yk ); @y yk (30)
0d; a; 0; p; q; with the added simplification that h(. . .,yk . . .) 0/0 at the primal solution, i.e. @h h (. . . ; y . . .)0 (y(yk ); @y yk (31)
min
ISE(V ; R; yf);
s:t: g1 (V ; R; yf)0 0; g2 (V ; R; yf)00; yf (yf 0 0;

k
0d; a; 0; p; q: Similarly, the terms in Eq. (6) that pertain to the inequality constraints become:
T lk g (. . . ; y . . .)
(35)
0lk g (. . . ; yk . . .)'lk 0p; q;
@g @y
(y(yk );
yk
(32)
where yf is a vector consisting of thirty elements yfi , i 0/ 1,. . ., 30, and the solution of Eq. (35) is denoted by k k V ; R and yf k : When applying Algorithm I, the k th master problem Eq. (16) for this example is: min h;
yf ;h
due to with the simplification that the complementarity optimality conditions of the primal problem. Hence: @g T T lk g (. . . ; y . . .) 0lk (y(yk ); (33) @y yk
T lk /g(. . .,yk . . .)0/0
s:t: h EISE(V ; R ; yf k )'vk (yf k (yf);
(36)
where the Lagrange multipliers vk satisfy the Karush/ Kuhn/Tucker (KKT) optimality conditions of Eq. (35): @(ISE) @V
0p; q: Substituting Eqs. (29) and (31) and Eq. (33) in Eq. (6) gives: L (u ; d k
k k
k k ; y)0J(xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; yk ; tf ) d a
@(ISE) @g @g 'lk 1 'lk 2 00; 1 2 @R k @R k @R k @(ISE) @(yf) 'lk 1 @g1 @(yf) k
@J ' ' @y yk
T 'jk d;f
@hd @ha T 'jk a;f @y tf ;yk @y tf ;yk
j g j

tf t0
nk d
@hd @y
dt'
yk
tf t0
nk a
@ha @y
dt
yk
j j j
'lk 1
@g1 @V
j j
'lk 2
@g2 @V
j j
00;
k
(37) (38) (39)
'lk 2
@g2
@(yf) k
'vk 00:
'jk d;0
@hd @y
'jk a;0
@ha @y
t0 ;yk
'kk
t0 ;yk
@h0 @y
Now assume that at the solution of the primal problem Eq. (35), inequality constraint g1 0/0 is inactive but g2 0/0 is active. In this case, lk 0/0; eliminating lk 1 2 from Eq. (37) and Eq. (39) then gives:
yk
T 'mk p
@hp @y
T 'lk p
@gp @y
yk
@(ISE) @g2 @(ISE) @V @(yf) k v 0( ' jk : @g2 @(yf) k @V
(40)
yk
@hq @gq T T 'mk 'lk (y(yk ); q q @y yk @y yk
(34)
Substituting Eq. (40) in Eq. (36) thus gives the following master problem for Algorithm I: min h;
yf;h
which is identical to Eq. (25) as required.
666
k k s:t: hE ISE(V ; R ; yf k ) @(ISE) (yf ( yf k ) ' @(yf) k @(ISE) @g2 @V @(yf) ( jk (yf ( yf k ): @g2 @V
min h;
yf
s:t: h EISE(V ; R ; yf k ) (41) ' @(ISE) @(yf) @(ISE) ( @V
(yf ( yf k )
k
@g2 @(yf) jk (yf ( yf k )
When applying the OA-based, Algorithm II, the k th master problem Eq. (21) for this example is: min h;
yf;V ;R;yf
@g2 @V
k k s:t: hE ISE(V ; R ; yf k ) ' @(ISE) @(ISE) k k (V ( V )' (R( R ) @V k @R k
@(ISE) @g2 @(ISE) @V @R ' ( jk @g2 @R k @V

(R( R ):
k
(44)
@(ISE) ' (yf ( yf k ); @(yf) k 0E g1 (V ; R ; yf k ) ' @g1 k (V ( V ) @V k

k k
However, the KKT optimality conditions Eq. (37) and Eq. (38) with lk 0/0 imply that: 1 @(ISE) @g2 @(ISE) @V @R ( jk 00: @g2 @R k
@g @g k ' 1 (R( R )' 1 (yf ( yf k ); @R k @(yf) k k k 0E g2 (V ; R ; yf k ) ' ' @g2 @V @g2 @R
j j j j
(45)
@V Thus, the master problem for Algorithm II becomes: min h; k

yf ;h
(V ( V )
k
(R( R )'
k
@g2
@(yf) k
(yf ( yf ); ' (42)
s:t: h EISE(V ; R ; yf k ) @(ISE) (yf ( yf k ) @(yf) k
00 yf (yf:
Now consider the case where, at the solution of the master problem above, the linearised constraint associated with g2 is active but the constraint associated with g1 remains inactive. In this case, the latter constraint can be ignored, while the equality of the former constraint implies that: k (V ( V ) ( 0 @g2 @(yf)
@(ISE) @g2 @V @(yf) ( jk (yf ( yf k ); @g2 @V
(46)
(yf ( yf k ) (
k
@g2 @R
k (R ( R ) : (43)
@g2 @V
Utilising Eq. (43) in Eq. (42) thus gives the following master problem for Algorithm II:
which is identical to the master problem for Algorithm I, Eq. (41). Hence, if g2 is active at the primal solution and is still the only active inequality constraint at the solution of the master problem, Algorithm II will give an identical solution to the master problem of Algorithm I. However, if any inequalities become active at the master solution that were inactive at the primal solution, then the master problem of Algorithm II will give a tighter lower bound than that of Algorithm I. This behaviour is indeed observed for Example 1, as shown in Table 4.
V. Bansal et al. / Computers and Chemical Engineering 27 (2003) 647 /668 Table 4 Comparison of the performance of the master problems of algorithms I and II for example 1 Iteration number Feed tray Algorithm I Algorithm II Lower bound, LB Algorithm I Algorithm II Active Inequalities, gk j Algorithm I Algorithm II LB from Algorithm I 0/LB from Algorithm II? 1 2 3 4
667
26 26 0.1576 0.1576 g1 1 g1 1 Yes
24 24 0.1813 0.1813 g2 1 g2 1 Yes
25 26 0.1815 0.1815 g2 1 g2 1 Yes
23 23 0.1848 0.2093 g3 1 g1, g2 1 1 No
References
Abel, O., & Marquardt, W. (2000). Scenario-integrated modeling and optimization of dynamic systems. American Institute of Chemical Engineering Journal 46 (4), 803 /823. Allgor, R. J., & Barton, P. I. (1999). Mixed-integer dynamic optimization I: problem formulation. Computers and Chemical Engineering 23 (4 /5), 567 /584. Androulakis, I. P. (2000). Kinetic mechanism reduction based on an integer programming approach. American Institute of Chemical Engineering Journal 46 (2), 361 /371. Avraam, M. P., Shah, N., & Pantelides, C. C. (1998). Modelling and optimisation of general hybrid systems in the continuous time domain. Computers and Chemical Engineering 22 (SS), S221 /S228. Avraam, M. P., Shah, N., & Pantelides, C. C. (1999). A decomposition algorithm for the optimisation of hybrid dynamic processes. Computers and Chemical Engineering 23 (SS), S451 /S454. Bagajewicz, M. J., & Manousiouthakis, V. (1991). On the generalized Benders decomposition. Computers and Chemical Engineering 15 (10), 691 /700. Balakrishna, S., & Biegler, L. T. (1993). A unied approach for the simultaneous synthesis of reaction, energy, and separation systems. Industrial Engineering and Chemical Research 32 (7), 1372 /1382. Bansal, V. (2000). Analysis, design and control optimization of process systems under uncertainty . PhD thesis. Imperial College, University of London. Bansal, V., Pistikopoulos, E. N., & Perkins, J. D. (2002). A case study in simultaneous design and control using rigorous, mixed-integer dynamic optimization models. Industrial Engineering and Chemical Research 41 (4), 760 /778. Bansal, V., Ross, R., Perkins, J. D., & Pistikopoulos, E. N. (2000). The interactions of design and control: double-effect distillation. Journal of Process Control 10 (2 /3), 219 /227. Bansal, V., Ross, R., Pistikopoulos, E. N., & Perkins, J. D. (2002). An industrial case study in simultaneous design and control using mixed-integer dynamic optimization. In: ESCAPE-12 Proceedings. Computer-Aided Chemical Engineering, Vol. 10 . Elsevier. Barton, P. I., Allgor, R. J., Feehery, W. F., & Galan, S. (1998). Dynamic optimization in a discontinuous world. Industrial Engineering and Chemical Research 37 (3), 966 /981. Biegler, L. T., & Sentoni, G. B. (2000). Efcient formulation and solution of nonlinear model predictive control problem. Latin American Applied Research 30 (4), 315 /324. Bloss, K. F., Biegler, L. T., & Schiesser, W. E. (1999). Dynamic process optimization through adjoint formulations and constraint
aggregation. Industrial Engineering and Chemical Research 38 (2), 421 /432. Brooke, A., Kendrick, D., & Meeraus, A. (1992). GAMS Release 2.25: a users guide . San Francisco: The Scientic Press. Bryson, A. E., & Ho, Y. (1975). Applied optimal control . New York: Taylor & Francis. Cervantes, A., & Biegler, L. T. (1998). Large-scale DAE optimization using a simultaneous NLP formulation. American Institute of Chemical Engineering Journal 44 (5), 1038 /1050. Cervantes, A., Tonelli, S., Brandolin, A., Bandoni, A., & Biegler, L. T. (2000). Large-scale dynamic optimization of a low density polyethylene plant. Computers and Chemical Engineering 24 (2 /7), 983 /989. Charalambides, M. S., Shah, N., & Pantelides, C. C. (1995). Synthesis of batch reaction/distillation process using detailed dynamic models. Computers and Chemical Engineering 19 (SS), S167 /S174. Dimitriadis, V. D., & Pistikopoulos, E. N. (1995). Flexibility analysis of dynamic systems. Industrial Engineering and Chemical Research 34 (12), 4451 /4462. Duran, M. A. (1984). A mixed-integer nonlinear programming approach for the systematic synthesis of engineering systems . PhD thesis. Carnegie-Mellon University. Duran, M. A., & Grossmann, I. E. (1986). An outer-approximation algorithm for a class of mixed-integer nonlinear programs. Mathematical Progress 36 (3), 307 /339. Fletcher, R., & Leyffer, S. (1994). Solving mixed integer nonlinear programs by outer approximation. Mathematical Progress 66 , 327 /349. Floudas, C. A. (1995). Nonlinear and mixed-integer optimization . New York: Oxford University Press. Floudas, C. A., & Visweswaran, V. (1990). A global optimization algorithm (GOP) for certain classes of nonconvex NLPs-I. Theory of Computers and Chemical Engineering 14 (12), 1397 /1417. Furlonge, H. I., Pantelides, C. C., & Sorensen, E. (1999). Optimal operation of multivessel batch distillation columns. American Institute of Chemical Engineering Journal 45 (4), 781 /801. Geoffrion, A. M. (1972). Generalized benders decomposition. Journal of Optimization Theory and Application 10 (4), 237 /260. Georgiadis, M. C., Rotstein, G. E., & Macchietto, S. M. (1998). Optimal design and operation of heat exchangers under milk fouling. American Institute of Chemical Engineering Journal 44 (9), 2099 /2111. Huang, Y. J., Reklaitis, G. V., & Venkatasubramanian, V. (2000). Dynamic optimization based fault accommodation. Computers and Chemical Engineering 24 (2 /7), 439 /444. Ishikawa, T., Natori, Y., Liberis, L., & Pantelides, C. C. (1997). Modeling and optimisation of an industrial batch process for the production of dioctyl phthalate. Computers and Chemical Engineering 21 (SS), S1239 /S1244. Lovik, I., Hillestad, M., & Hertzberg, T. (1998). Long term dynamic optimization of a catalytic reactor system. Computers and Chemical Engineering 22 (SS), S707 /S710. Mcauley, K. B., & Macgregor, J. F. (1992). Optimal grade transitions in a gas-phase polyethylene reactor. American Institute of Chemical Engineering Journal 38 (10), 1564 /1576. Mohideen, M. J., Perkins, J. D., & Pistikopoulos, E. N. (1996). Optimal design of dynamic systems under uncertainty. American Institute of Chemical Engineering Journal 42 (8), 2251 /2272. Mohideen, M. J., Perkins, J. D., & Pistikopoulos, E. N. (1997). Towards an efcient numerical procedure for mixed integer optimal control. Computers and Chemical Engineering 21 (SS), S457 /S462. Narraway, L. T. (1992). Selection of Process Control Structure Based on Economics . PhD thesis. Imperial College, University of London. Nath, R., & Alzein, Z. (2000). On-line dynamic optimization of olens plants. Computers and Chemical Engineering 24 (2 /7), 533 /538.
668
V. Bansal et al. / Computers and Chemical Engineering 27 (2003) 647 /668 Schweiger, C. A., Rojnuckarin, A., & Floudas, C.A. (1997). MINOPT: users guide . Princeton University. Software Version 2.0. Sharif, M., Shah, N., & Pantelides, C. C. (1998). On the design of multicomponent batch distillation columns. Computers and Chemical Engineering 22 (SS), S69 /S76. Sullivan, G. R., & Sargent, R. W. H. (1979). Development of feed changeover policies for renery distillation units. Industrial Engineering and Chemical Process Design and Development 18 , 113 / 124. Vassiliadis, V. (1993). Computational solution of dynamic optimization problems with general differential-algebraic constraints . PhD thesis. Imperial College, University of London. Vassiliadis, V. S., Sargent, R. W. H., & Pantelides, C. C. (1994a). Solution of a class of multistage dynamic optimization problems. Part I-problems with path constraints. Industrial Engineering and Chemical Research 33 (9), 2111 /2122. Vassiliadis, V. S., Sargent, R. W. H., & Pantelides, C. C. (1994b). Solution of a class of multistage dynamic optimization problems. Part II-problems with path constraints. Industrial Engineering and Chemical Research 33 (9), 2123 /2133. Viswanathan, J., & Grossmann, I. E. (1990). A combined penalty function and outer-approximation method for MINLP optimization. Computers and Chemical Engineering 14 (7), 769 /780. Viswanathan, S., Shah, N., & Venkatasubramanian, V. (2000). A hybrid strategy for batch process hazards analysis. Computers and Chemical Engineering 24 (2 /7), 545 /549. White, V., Perkins, J. D., & Espie, D. M. (1996). Swithchability analysis. Computers and Chemical Engineering 20 (4), 469 /474.
Pantelides, C.C. (1996). Process Modelling in the 1990s. In: Fifth world congress of chemical engineering . San Diego, California. Process Systems Enterprise Ltd. (2000). gPROMS advanced users guide. Available from: http://www.psenterprise.com. Pytlak, R. (1998). Runge /Kutta based procedure for optimal control of differential /algebraic equations. Journal of Optimization Theory and Application 97 (3), 675 /705. Rosen, O., & Luus, R. (1991). Evaluation of gradients for piecewise constant optimal control. Computers and Chemical Engineering 15 (4), 273 /281. Ross, R., Bansal, V., Perkins, J. D., Pistikopoulos, E. N., Koot, G. L. M., & van Schijndel, J. M. G. (1999). Optimal design and control of an industrial distillation system. Computers and Chemical Engineering 23 (SS), S875 /S878. Ross, R., Perkins, J. D., & Pistikopoulos, E. N. (1998). A mixedinteger dynamic optimization approach to simultaneous design and control. Paper 220e, AIChE annual meeting . Miami Beach, Florida. Sahinidis, N. V., & Grossmann, I. E. (1991). Convergence properties of generalized Benders decomposition. Computers and Chemical Engineering 15 (7), 481 /491. Samsatli, N. J., Papageorgiou, L. G., & Shah, N. (1998). Robustness metrics for dynamic optimization models under parameter uncertainty. American Institute of Chemical Engineering Journal 44 (9), 1993 /2006. Schweiger, C. A., & Floudas, C. A. (1997). Interaction of design and control: optimization with dynamic models. In W. W. Hager & P. M. Paraldos (Eds.), Optimal control: theory, algorithms and applications (pp. 388 /435). Kluwer Academic Publishers BV.

Science

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Science

Uploaded by

Copyright:

Available Formats

Computers and Chemical Engineering 27 (2003) 647 /668 www.elsevier.

New algorithms for mixed-integer dynamic optimization

V. Bansal et al. / Computers and Chemical Engineering 27 (2003) 647 /668

2. Preliminaries A MIDO problem can be formulated as:

min J(xd (tf ); xd (tf ); xa (tf ); u(tf ); d; y; tf );

V. Bansal et al. / Computers and Chemical Engineering 27 (2003) 647 /668

V. Bansal et al. / Computers and Chemical Engineering 27 (2003) 647 /668

    @hp T k @gp T k ' mp ' lp 00; @xd @xd

T k k nk (t)hd (xd (t); xk (t); xk (t); uk (t); d ; y; t)dt d a d

nk (t)ha (xk (t); xk (t); uk (t); d ; y; t)dt d a a

V. Bansal et al. / Computers and Chemical Engineering 27 (2003) 647 /668

'kk h0 (xd (t0 ); xk (t0 ); xk (t0 ); uk (t0 ); d ; y; t0 ) d a

'lk gp (xd (ti ); xk (ti ); xk (ti ); uk (ti ); d ; y; ti ) d a p 'mk hq (d ; y)'lk gq (d ; y); q q

min J(xd (tf ); xd (tf ); xa (tf ); u(tf ); d; y; tf );

T k k nk (t)hd (xd (t); xk (t); xk (t); uk (t); d ; yk ; t)dt d a d

V. Bansal et al. / Computers and Chemical Engineering 27 (2003) 647 /668

nk (t)ha (xk (t); xk (t); uk (t); d ; yk ; t)dt d a a

T k k 'jk hd (xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; yk ; tf ) d a d;f

k h0 (xk (t0 ); xk (t0 ); xk (t0 ); uk (t0 ); d ; yk ; t0 ) d d a

T k k 'mk hp (xd (ti ); xk (ti ); xk (ti ); uk (ti ); d ; yk ; ti ) d a p

T k k 'lk gp (xd (ti ); xk (ti ); xk (ti ); uk (ti ); d ; yk ; ti ) d a p

'mk hq (d ; yk )'lk gq (d ; yk ) q q 'vk (yk ( y);

k s:t: Lk (uk ; d ; y)0 h;

s:t: hd (xd (t); xd (t); xa (t); u(t); d; y; t)00; t  [t0 ; tf ];

V. Bansal et al. / Computers and Chemical Engineering 27 (2003) 647 /668

k k d a s:t: J(xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; y; tf ) '

T k k nk (t)hd (xd (t); xk (t); xk (t); uk (t); d ; y; t)dt d a d

nk (t)ha (xk (t); xk (t); uk (t); d ; y; t)dt d a a

T k k 'jk hd (xd (tf ); xk (tf ); xk (tf ); uk (tf ); d ; y; tf ) d a d;f

0h; k  Kfeas ; vk (yk ( y)00;

T k k 'lk gp (xd (ti ); xk (ti ); xk (ti ); uk (ti ); d ; y; ti ) d a p

'mk hq (d ; y)'lk gq (d ; y) q q k  Kinfeas ; (14) 0h: (15)

V. Bansal et al. / Computers and Chemical Engineering 27 (2003) 647 /668

J(xd (tf ); xd (tf ); xa (tf ); u(tf ); d; y; uy (tf ); tf );

J(xd (tf ); xd (tf ); xa (tf ); u(tf ); d; y; uy (tf ); tf );

i 01; . . . ; N; hq (d; y)00; gq (d; y)00;

V. Bansal et al. / Computers and Chemical Engineering 27 (2003) 647 /668

hy 0y( yk 00; huy (j)0uy (j)( uk (j)00; y j 01; . . . ; Nuy : (19)

s:t: h EJ ' '

k (d( d )' (uj ( uk ); j

    @gp k )' @gp (y(yk ) (d( d @d k @y k X@gp  ' (uj ( uk ); j @uj k j

J 'v (y ( y)' k  Kfeas ;

[vky (j)]T [uk (j)( uy (j)]0h; y u

      @hq @hp k 0 ETk hk ' (d( d )' (y(yk ) ; p q @d k @y k

Nuy X [vky (j)]T [uk (j)( uy (j)] 00; y v (y ( y)' u kT k j01

    @gq @hp k (d( d )' (y(yk ); @d k @y k

0 0y( y; k  Kfeas ; (21)

V. Bansal et al. / Computers and Chemical Engineering 27 (2003) 647 /668

V. Bansal et al. / Computers and Chemical Engineering 27 (2003) 647 /668

(5) Initial conditions: ISE(t0 0)0 0;

0 (xb (x+ )2 '(xN'1 (x+ )2 ; b N'1

yfi 01; i 01; . . . ; 9; i 010; . . . ; N;

yfi  f0; 1g;

Fig. 1. Superstructure for Example 1.

fyf ;yr;Dcol ;Kv ;Kr ;tv tr ;x+ ;x+ g b N'1

07756Vmean '1891:125'996:3D2 col

660 Table 2 (Continued ) Iteration number Feed tray V , R (kmol/min) LB UB(/LB0/0? 1

V. Bansal et al. / Computers and Chemical Engineering 27 (2003) 647 /668

2 24 1.5179, 0.9777 0.1813 No

3 25 1.6375, 1.0977 0.1815 No

4 23 1.3856, 0.8440 0.2093 Yes: STOP

26 1.7488, 1.2072 0.1576 No

@hp T k @gp T k ' mp ' lp 00; @xd @xd

s:t: hd (xd (t); xd (t); xa (t); u(t); d; y; t)00; t [t0 ; tf ];

0h; k Kfeas ; vk (yk ( y)00;

'mk hq (d ; y)'lk gq (d ; y) q q k Kinfeas ; (14) 0h: (15)

@gp k )' @gp (y(yk ) (d( d @d k @y k X@gp ' (uj ( uk ); j @uj k j

J 'v (y ( y)' k Kfeas ;

@hq @hp k 0 ETk hk ' (d( d )' (y(yk ) ; p q @d k @y k

@gq @hp k (d( d )' (y(yk ); @d k @y k

0 0y( y; k Kfeas ; (21)

yfi f0; 1g;

N X '1494:45Dcol 6'0:76 i yri

dxN'1 0 V (yN (xN'1 ); dt 0:0014134 Dcol 2

(3) Overall balances: dVmean V 0 ; tf dt s.t. (2) Component balances: 10m X N

yfi ; yri f0; 1g;

(9) Sinusoidal disturbances: 2pt ; F 01'0:2sin 180 zf 00:45'0:09sin 2pt 60 ;

s:t: h ECostp ' k Kfeas ;

35 X '1494:45Dcol 6'0:76 i yri

@hp @gp T 'lk p @y yk @y yk

@hq @gq T T 'mk 'lk (y(yk ): q q @y yk @y yk

T @hd ' @y T @hd ' @y

T T T @hp @gp @hq k k ' mp ' lp ' @y @y @y yk yk T @gq ' lk 'Ivk : q @y yk

0lk g (. . . ; yk . . .)'lk 0p; q;

T lk /g(. . .,yk . . .)0/0

@hd @ha T 'jk a;f @y tf ;yk @y tf ;yk

@hq @gq T T 'mk 'lk (y(yk ); q q @y yk @y yk