R&D Notes Steady-state multiplicities in reactive distillation: stage-by-stage calculation revisited

Ali Baharev , Arnold Neumaier March 28, 2012

Fakultt fr Mathematik, Universitt Wien a u a Nordbergstrae 15, A-1090 Wien, Austria email: ali.baharev@gmail.com, Arnold.Neumaier@univie.ac.at (*) Author to whom all correspondence should be addressed. Abstract. Bifurcation diagrams are computed for a reactive distillation column producing ethylene glycol from ethylene oxide and water. The column can have as many as 7 steady states. The rigorous steady-state model of a reactive distillation column, a large-scale system of nonlinear equations, is dicult to solve with current nonlinear solvers. The stage-by-stage method, which reduces this high-dimensional problem to a low-dimensional one, is well-known since the 1930s (Lewis & Matheson [5]). Unfortunately, the nal, lowdimensional system is so nonlinear, with an extreme sensitivity to the initial estimates, that solving it is very dicult if not impossible. For this reason, this method is not implemented in modern process simulators. Here a proof-of-concept method is presented that remedies the numerical diculties of the stage-by-stage calculation. The proposed algorithm does not require initial column prole estimates. Convergence is guaranteed, assuming the bubble point calculation succeeds on the stages. Keywords: multiple steady states, tearing methods, homotopy continuation, bifurcation, ethylene glycol

Essential dimension. We dene the essential dimension of a system of equations as the minimal number d of variables that need to be xed so that the remaining variables can be calculated solving only small nonlinear systems sequentially. Thus the stage-by-stage calculation results in a system of d equations for those d variables that were used to start the calculation. Typically, the essential dimension is far smaller than the total number of variables, which would make the stage-by-stage method attractive computationally were it not for the severe numerical instability that results. For problems with essential dimension d = 1, we show how to implement the stage-bystage method in a numerically stable way. The successful computation of the bifurcation diagrams for the above problem serves as proof of concept for the approach. Signicant eorts were made to compute these bifurcation diagram with the state of the art solver IPOPT (Wachter & Biegler [6]) and with various problem reformulations (Biegler [1], Ch. 7). These attempts failed, perhaps due to scaling issues and ill-conditioning. A coupled reactor-column system. The model is taken from Yi & Luyben [7] (pp. 3438). The aim of this example is to demonstrate the fundamental aw of the traditional stage-by-stage method. The stage-by-stage method is still able to solve the problem but shows sensitivity problems. These can become extremely severe in more complex problems, causing the conventional stage-by-stage method to fail. The process consists of a rectifying column sitting above a continuous stirred tank reactor (CSTR). Vapor from the reactor passes directly into the column. A reversible, rst-order reaction A B occurs only in the liquid phase inside the reactor. In this binary system, the product is more volatile than the reactant. The reactor hold-up (VR ) is considered constant, the feed ow (F0 ) controls the level (hence VR ) in the reactor. The reactor temperature (TR ) is not controlled. The heat input to the reactor (QR ) and the reux ow rate (L) are specied. The distillate ow rate (D) and composition (xD ) are sought. This corresponds to the inventory structure S100 (p. 34). The rigorous model, coded in AMPL modeling language (Fourer et al. [4]), is included in the online supplementary material. The problem has two steady-state solutions. The model has 25 variables and 25 equations, but its essential dimension is one. The only

variable needed to start the stage-by-stage calculation is TR . After TR is guessed, a univariate Newton method is executed to nd the composition in the reactor, z. (Alternatively, z could be guessed and TR determined by the Newton method.) Then the stage-by-stage calculation can be started. The nal equation is xN +1 = yN , where xN +1 is the composition of the liquid leaving, and yN is the composition of the vapor entering the total condenser.
xN +1 yN 1.0

0.5

0.0 140 -0.5 142 144 146 148 150 TR

-1.0

Figure 1: Coupled reactor-column system, residual in the nal equation as a function of the reactor temperature TR . The function is singular at TR 144.87 due to the denominator becoming zero in the vapor-liquid equilibrium condition.

Narrow region of attraction. In Figure 1, xN +1 yN is plotted against TR . The sensitivity of the nal equation to the initial estimate is apparent. If a strict descending local search procedure is used to minimize the merit function (e.g. (xN +1 yN )2 or |xN +1 yN |), the solution at TR = 144.80 is found only if the procedure is started from [144.32, 144.87]. The latter interval is called the region of attraction. The narrow region of attraction illustrates well why the stage-by-stage method is prone to fail. If the proposed method is executed starting with a grid {140, 141, . . . , 150} of initial estimates for TR , it nds both solutions. In contrast, a multistart method with strict descending local search procedure works only if the initial grid has points in the region of attraction of both solutions. The solution at TR = 140.7 has a wide region of attraction, and the multistart method nds it ve times. However, the initial manifold has no point in the region of attraction of the solution at TR = 144.8, hence misses this solution. More dicult problems

of the same kind have at least one solution with an extremely narrow regions of attraction, so that extremely ne grids and multiple precision arithmetic would be needed to nd the solution. Outline of the proposed method. The algorithm is presented only at a conceptual level, due to the space limitations. For implementation details, the reader is referred to the C++ source code, available in the online supplementary material. Discretized 1D manifolds. Let the variable TR have the following bounds 140 TR 150. The orderer set of points {140, 141, 142, . . . , 150} and linear interpolation between the neighboring points dene a piecewise linear (PL) manifold. This manifold covers the variable TR in the region of interest. Forward sweep. The proposed method starts with the PL manifold covering the initial variable. This PL approximation to manifold is propagated through the column, stage-bystage. The propagation is achieved by propagating the breakpoints of the PL manifold. At each stage, we save the corresponding input-output points approximating the input-output manifold map. Figure 2 shows an example. The online supplementary material gives further explanatory color gures. After each stage, the breakpoints of the PL approximation are redistributed (detailed later how) so that the representation of the manifold remains fairly good, see Figure 2. The propagation is continued with the new set of points. The input-output map of the stage is not changed by the redistribution of the points. The stagewise redistribution of the points also means that the output manifold of a stage is only available in terms of its input manifold but it does not depend directly on any other previous variables. This is the key point to achieve numerical stability, as we no longer have to deal with the extreme sensitivities. Finally, we get a piecewise linear representation of the nal equation to be solved. It is solved with linear interpolation. Backward sweep. The price of the stage-wise redistribution of the points is that we have to pass through input-output manifold maps of the stages, in the reverse-order as saved, and apply linear interpolation to get the values of the intermediate and the initial variables.

0.8 0.6 0.4 0.2 0.0 0.0 x2 1.0 0.8 0.2 0.4 x1 (2c) 0.6 0.8 1.0

arc length, output manifold

x2 1.0

(2a)

1.0 0.8 0.6 0.4 0.2 0.0 0.0

(2b)

0.2 0.4 0.6 0.8 arc length, input manifold

(2d)

1.0

0.9999 0.999 0.99 x3 at stage 6

0.99970 0.99965

0.6 0.4 0.2 0.0 0.0 0.2 0.4 x1 0.6 0.8 1.0

0.9 0.5 0.1 0.01 0.001 0.0001 0 12 5 10 20 50 100

0.99960 3.83 3.85

Figure 2: (ac) The map between the input and output manifolds at stage 1; (d) bifurcation diagram. Mole fraction in the liquid phase is denoted by x. Components: (1) ethylene oxide, (2) water, (3) ethylene glycol. The bifurcation parameter is proportional to the hold-up volume on the reactive stages. Redistribution of points. Resampling distinguishes the proposed method from the stageby-stage multistart methods, which would fail for the same reason why the original stage-bystage method itself fails. The manifold representation is improved after each stage in order to avoid the numerical diculties that the stage-by-stage method has. Such improvements are possible (only) because we propagate the entire manifold stage-by-stage and not just a single point. (a) If any two neighboring points become too distant then new points have to be inserted between them. Otherwise the representation of the manifold can become very poor on the next stage as we have no information in the region between the distant points.

(b) If some of the points become too close at a fairly linear part of the manifold then some of these points have to be removed. These points barely contribute to the representation of the manifold but the accumulation of such points can cause a serious degradation of performance. (c) Infeasible points can make the function evaluations fail. Infeasible parts of the manifold have to be discarded, and points on the boundary to infeasibility may have to be introduced. Computing bifurcation diagrams. In many studies one is interested in the dependence of the characteristics on a design parameter (the bifurcation parameter) that can be varied, resulting in bifurcation diagrams. Conventionally, continuation methods track a single solution as the bifurcation parameter is varied. Turning points and bifurcation points in the bifurcation diagram require special attention, as simple continuation becomes singular there. All solutions for a xed bifurcation parameter are available only after the bifurcation diagram is completed, and only under the assumption that all connected components of the diagram have been found. This assumption is dicult to check, and it is easy to miss isolas in the bifurcation diagram. The proposed method constructs the bifurcation diagram in an orthogonal fashion. In one step, all solutions are calculated for a xed value of the bifurcation parameter. Turning points and isolas require no special attention. The fact that consistent curves are computed serves as an important consistency check on the quality of the method. Reactive distillation column for manufacturing ethylene glycol. The rigorous model of the reactive distillation column is taken from Ciric & Miao [3], corresponding to the costoptimal column of Ciric & Gu [2]. Ethylene glycol is produced from ethylene oxide and water via the reactions C2 H4 O + H2 O C2 H6 O2 C2 H4 O + C2 H6 O2 C4 H10 O3 . Ethylene glycol is the major product of these reactions, and diethylene glycol is a waste byproduct. The second reaction is neglected to make the problem essentially 1-dimensional. As shown in subsection 7.3. Multiple reactions of Ciric & Miao [3], this does not change the bifurcation diagram qualitatively.

(1)

The column has 10 stages, numbered from the bottom to the top. Reaction and separation occur on stages 5 through 10 in the column, and separation without chemical reaction occurs on stages 1 through 4. Water and ethylene oxide are fed to the column in a strict 1:1 ratio. The water is fed onto the top stage, and the ethylene oxide is distributed among stages 6-10. The column produces no distillate products. The boil-up ratio is specied. The rigorous model, coded in AMPL modeling language (Fourer et al. [4]), is included in the online supplementary material. The model has 70 variables and 70 equations. Again, the essential dimension equals one. The bifurcation parameter considered here is proportional to the hold-up volume on the reactive stages. Bifurcation diagrams are given in Figure 2d. The column has 7 steady-states in a particular region of . The bifurcation diagrams match the results of Ciric & Miao [3] qualitatively but there are quantitative dierences. (Honest eorts have been made to resolve this. Unfortunately the reason remains unknown, perhaps some of the model parameters in Ciric & Miao [3] contain a misprint.) Future work. The assumption behind the current implementation of the method is that the individual stages do not show strong sensitivity. Unfortunately, this assumption turned out to be false. An unreasonably large number of points (10000) is currently needed to maintain a good representation of the manifold while propagated through such an ill-conditioned stage. This issue is the rst to be resolved by adaptively selecting the points to be propagated. It requires signicant changes in the bookkeeping logic which is already complicated. The current implementation uses linear interpolation when it passes through the input-output manifold maps of the stages. Linear interpolation cannot resolve nonlinear behavior: if the individual stages have multiple steady states, it will be unnoticed. This will be xed by applying a global solver instead of linear interpolation. Although the idea of the algorithm is easy to present, its implementation is complicated even for 1-dimensional problems due to the bookkeeping logic. The current implementation is intentionally restricted to 1-dimensional problems. It makes the implementation easier and the visualization of the results simpler (2D plots). The method will be extended to higher essential dimensions, for example multiple columns with recycles.

Acknowledgements. The research was funded by the Austrian Science Fund (FWF): P23554.

References
[1] Lorenz T. Biegler. Nonlinear programming: concepts, algorithms, and applications to chemical processes. SIAM, 2010. [2] Amy R. Ciric and Deyao Gu. Synthesis of nonequilibrium reactive distillation processes by MINLP optimization. AIChE Journal, 40:14791487, 1994. [3] Amy R. Ciric and Peizhi Miao. Steady state multiplicities in an ethylene glycol reactive distillation column. Ind. Eng. Chem. Res., 33:27382748, 1994. [4] Robert Fourer, David M. Gay, and Brian Wilson Kernighan. AMPL: A Modeling Language for Mathematical Programming. Brooks/Cole USA, 2003. [5] W. K. Lewis and G. L. Matheson. Studies in distillation. Industrial and Engineering Chemistry, 24:494498, 1932. [6] Andreas Wchter and Lorenz T. Biegler. On the implementation of an interior-point lter a line-search algorithm for large-scale nonlinear programming. Mathematical Programming, 106:2557, 2006. [7] Chang K. Yi and William L. Luyben. Design and control of coupled reactor/column systemsPart 1. A binary coupled reactor/rectier system. Computers & Chemical Engineering, 21(1):2546, 1996.